FIM-CNN for detection of live cells and / or particulate impurities

Abstract

The disclosure presented herein provides a method for quantifying viable cells and particulate cell impurities in a cell-based product sample. The method is implemented on a convolutional neural network (CNN) that learns to classify flow-imaging microscopy (FIM) images. Training of the CNN is accomplished by using a training set of classified images of viable cells and different types of impurities.

Description

[Technical field]

[0001] The disclosure presented herein provides a method for quantifying viable cells and particulate cell impurities in a cell-based product sample. The method is implemented on a convolutional neural network (CNN) that learns to classify flow-imaging microscopy (FIM) images. Training of the CNN is accomplished by using a training set of classified images of viable cells and different types of impurities. [Background technology]

[0002] The number of cell-based medicinal products (CBMPs) that have entered clinical trials and been approved by major regulatory agencies for commercial use is consistently increasing every year. Despite the encouraging clinical data emerging from the use of these innovative therapeutic products, many challenges remain in the areas of manufacturing, formulation development and analytical characterization.

[0003] Most CBMPs consist of living cells that are inherently fragile and much more susceptible to adverse conditions compared to other biotherapeutics such as protein-based products. Nevertheless, regardless of the type of CBMP (e.g., autologous or allogeneic, genetically modified or non-modified), they often undergo multiple complex manufacturing steps before being administered to patients. The main processing steps that CBMPs undergo during their production include isolation of cells from healthy donors or patients, ex vivo cell manipulation, formulation, storage, and quality control prior to release and administration. In addition, transportation, freeze-thawing or manual manipulation of the CBMP occurs during each of the steps listed above. Clinical site-specific handling procedures also result in differences in product processing and administration. Such diverse and multi-step manufacturing and handling processes expose cells to a variety of intentional and / or accidental environmental stressors such as freeze-thawing, surface-related stress (i.e., mechanical stress), thermal and oxidative stress. Each of these types of stresses can result in accidental cell death. Therefore, testing of the effect of formulation parameters on cell stability under these stress conditions should be considered to better understand the product's sensitivity to the stressors involved. This is likely to contribute new insights that can be used to mitigate the potential risks of treatment failure and the occurrence of serious adverse effects due to poor product quality.

[0004] The introduction or formation of particulate impurities in CBMPs is one of the potential risks associated with the manufacture and handling of these formulations. Particles found in CBMPs can originate from either the process or the cell product itself. Process-related particulate impurities can include ineffectively removed antibody-coated magnetic beads used for cell activation and proliferation, viral vectors utilized in cell transduction steps, exogenous particles (glass, fiber, rubber), or leachates from primary containers. Sterile filtration of cell suspensions is not possible due to the inherent size of the cells. Thus, if their removal is not complete, undesirable micrometer-sized particles remain in the product. Cellular impurities, on the other hand, include materials derived from cells. These can be non-viable or non-therapeutic cells, as well as cell aggregates and cell debris. For example, necrosis induced in cells exposed to extreme physiological conditions (pyroptosis) leads to rupture of the plasma membrane, release of intracellular contents and formation of debris particles. In contrast to apoptotic (programmed) cell death, in which cellular components are packaged into vesicles and digested by appropriate caspases to facilitate clearance by the immune system, necrosis generates debris with strong inflammatory properties. One consequence from these debris particles could be an adverse immune response upon administration. Furthermore, the introduction of larger micrometer-sized particles into smaller blood vessels could cause tissue damage due to thromboembolism.

[0005] Forced degradation studies are generally included in the development of any pharmaceutical product as part of the establishment of analytical methods and formulation screening. In addition to real-time stability studies at intended storage conditions, forced degradation studies are applied to estimate the shelf life of, for example, protein-based biopharmaceuticals and to define appropriate storage and handling conditions. Furthermore, this type of study can simulate accidental exposure of a drug product to harmful conditions and aid in the evaluation of potential risks that may occur to the drug product throughout the product's lifecycle. These studies also help in the selection of appropriate excipients during formulation development, the objective of which is to determine the component mixture that achieves maximum stability of the active pharmaceutical ingredient. Although the manufacturability, critical quality attributes (CQAs), degradation pathways and product degradants in CBMPs are substantially different than, for example, protein-based therapeutics, the same concepts of forced degradation studies can be highly relevant to the process and product development of CBMPs.

[0006] Extensive testing of CBMPs prior to administration to patients is a regulatory mandate to ensure the product is safe, effective and of good quality. Testing parameters include cell identity, purity, activity and potency, but cells and non-cellular materials must also be identified and qualified. To this end, reliable analytical methods for product quality characterization and stability testing are required.

[0007] There remains a need for automated, label-free methods for quantifying impurities in cell-based products. Neural networks have been shown to be beneficial for classifying different types of images. However, their application for the analysis of flow imaging microscopy (FIM) of cell products involves several challenges, such as training neural networks on a substantial dataset consisting of precisely selected cell images from each population. As disclosed herein, obtaining such training datasets encompasses several difficulties, and datasets with mixed particle images result in a large number of errors in the classification of FIM-acquired cell images. Moreover, it is not predictable whether neural networks will be robust enough to be trained on datasets containing such a large number of misclassified images.

[0008] This application discloses flow imaging microscopy combined with convolutional neural networks (FIM-CNN) to establish a high-throughput and label-free method for quantification of live cells, necrotic cells, apoptotic cells, debris particles, magnetic beads, and adducts (single cells with bead(s) attached). Summary of the Invention

[0009] In one aspect, disclosed herein is a method for quantifying single live cells and / or particulate impurities in a cell-based product sample, the method comprising: a) providing a training set of classified flow imaging microscopy (FIM) images comprising single live cells, debris, and at least one of necrotic cells, apoptotic cells, doublet cells, magnetic beads, or adducts; b) training a convolutional neural network (CNN) with the classified images of step (a); c) capturing FIM images from the product sample; d) applying the trained CNN obtained in step (b) to the FIM images of step (c) to identify at least one of single live cells, necrotic cells, apoptotic cells, doublet cells, magnetic beads, adducts, and debris; and e) quantifying the single live cells and / or particulate impurities identified in step (d).

[0010] In some related aspects, the set of classified images includes: a single live cell, a necrotic cell, an apoptotic cell, a doublet cell, a magnetic bead, an adduct, and debris; a single live cell, a doublet cell, a magnetic bead, an adduct, and debris; a single live cell, a necrotic cell, and debris; or a single live cell, a necrotic cell, an apoptotic cell, and debris.

[0011] In some related embodiments, the classified image of step (a) is obtained by a method comprising: i) providing a cell suspension; ii) applying cell suspension conditions to induce necrosis, apoptosis, cell debris, appendages, or any combination thereof in the cell culture; iii) optionally labeling cells; iv) acquiring a FIM image of the cell culture; and v) classifying the FIM image.

[0012] In some related aspects, the culture conditions for inducing necrosis and / or inducing debris comprise subjecting the cell culture to stress conditions, the culture conditions for inducing apoptosis comprise incubating the cell culture with a pro-apoptotic reagent, the culture conditions for inducing adducts comprise incubating the cells with magnetic beads, or any combination thereof.

[0013] In some related embodiments, the stress conditions for inducing necrosis include shaking stress, freeze-thaw cycles, heat, incubation with ethanol, or any combination thereof. In some related embodiments, the pro-apoptotic agent includes staurosporine. In some related embodiments, the cells are incubated with the magnetic beads in about a 1:1 ratio.

[0014] In some related embodiments, the conditions to induce necrosis include incubating the cells in cell culture containing 10% ethanol at 55° C. for 90 minutes or at 37° C. for 90 minutes, the conditions to induce apoptosis include incubating the cells with 1 μM staurosporine at 37° C. for 3 hours, the conditions to induce debris include vortexing the cells in a 2 mL microfuge for 2 minutes or subjecting the cells to two freeze-thaw cycles from −140° C. to 37° C., and the conditions to induce adducts include incubating magnetic beads in a 1:1 ratio with the cells at 37° C. for 1.5 hours.

[0015] In some related embodiments, the necrotic cells from the training set include cells incubated in cell culture containing 10% ethanol for 90 minutes at 37° C. and cells incubated for 90 minutes at 55° C. In some related embodiments, the classified image of step (a) comprises i) obtaining a sample of the cell-based product, ii) optionally labeling cells of the sample of step (i), iii) obtaining a FIM image of the sample of step (i) or (ii), and iv) classifying the FIM image.

[0016] In some related aspects, the live cells from the training set include cells incubated in hypotonic medium, cells incubated in isotonic medium, and cells incubated in hypertonic medium.

[0017] In some related aspects, the CNN is pre-trained prior to step (b) using a pre-training set of images. In some related aspects, the CNN comprises a VGG-19 architecture. In some related aspects, during the training of step (b), only the last two fully connected layers of the CNN are active, or the first 10 layers of the CNN are frozen.

[0018] In some related embodiments, the training in step (b) comprises an Adam optimization algorithm or an SGD optimization algorithm. In some related embodiments, the cell product comprises T cells.

[0019] In some aspects, disclosed herein are methods for quantifying single live cells and / or particulate impurities in a cell-based product sample, the method comprising: a) providing a cell suspension comprising single live, necrotic, and / or apoptotic cells; b) simultaneously obtaining brightfield images and fluorescent signals from the cell suspension using an imaging flow cytometer, the fluorescent signals being used to assign a classification label to the brightfield image according to whether the brightfield image comprises live, necrotic, or apoptotic cells; c) training a convolutional neural network (CNN) with the brightfield images of step (b); d) capturing flow imaging microscopy (FIM) images from the product sample; e) identifying single live cells, debris, and at least one of necrotic cells, apoptotic cells, doublet cells, magnetic beads, or adducts by applying the trained CNN obtained in step (c) to the FIM images of step (d); and f) quantifying the single live cells and / or particulate impurities identified in step (e). [Brief description of the drawings]

[0020] [Figure 1] FIG. 1 shows the clustered morphological parameters used to distinguish five particle populations by using the output data from FlowCam software. The arrows with accompanying numbers indicate the order in which separation of particle populations was performed. [Diagram 2] Schematic of the convolutional neural network (CNN) workflow. First, a set of 4,000 to 4,500 images from each particle class is manually selected for retraining with the VGG-19 network. During retraining, the weight kernels in the last two layers of the network update the weight parameters and extract representative image descriptors based on the input data. Once retraining is complete, the network can be used to predict particle classes for new images. The output is given as the probability of an image being assigned to a specified class. [Diagram 3] FIG. 1 shows the recovery (concentration) of Dynabeads determined by using flow imaging microscopy with respect to the theoretical concentration of Dynabeads in three different preparation setups. [Figure 4] FIG. 1 shows the concentration of adducts measured by FIM-CNN. [Diagram 5] FIG. 1 shows images collected by using FlowCam with incorrect focus settings (top) and correct focus settings (bottom). [Figure 6] Confusion matrices obtained for the training results of the machine learning model. The image datasets included necrotic cells obtained from digitonin treatment (FIG. 6A), heat and ethanol treatment (FIG. 6B), heat and ethanol treatment + curation of the image datasets for each class using morphological parameters (FIG. 6C). [Figure 7A] FIG. 1 shows representative images of each population class obtained by using FlowCam. [Figure 7B] FIG. 1 shows representative images of each population class obtained by using FlowCam. [Figure 7C]Figure 7 shows representative images of each population class obtained by using FlowCam. All images have been enlarged for presentation purposes. Figure 7A shows the particle size distribution of a sample containing Jurkat cells at a concentration of 375,000 cells / mL. Figure 7B shows the particle size distribution of a sample containing Dynabeads at a concentration of 68,000 beads / mL. Figure 7C shows the particle size distribution of a sample containing Jurkat cells with Dynabeads. [Figure 8A] FIG. 13 shows classification of Dynabeads images by using morphological parameters output from FlowCam software. [Figure 8B] Classification of Dynabeads images by using morphological parameters output from FlowCam software. Figure 8A shows the concentration of Dynabeads measured at a range of Dynabeads concentrations suspended in RPMI buffer (DB control, grey) and cell suspension (DB+cells, black). Figure 8B shows the recovery (left y-axis) and coefficient of variation (right y-axis) of Dynabeads in RPMI buffer (grey) and cell suspension (black). Error bars represent the standard deviation of the mean of six replicates. The coefficient of variation (%) was calculated from six replicates. [Figure 9A] FIG. 1 shows the results of Dynabeads (DB) classification using CNN. [Figure 9B] Figure 9 shows the results of Dynabeads (DB) classification using CNN. Figure 9A shows the determination of the concentration of Dynabeads in a concentration series of Dynabeads suspended in RPMI medium (DB control, grey) and Jurkat cell suspension (DB+ cells, black). Figure 9B shows the recovery (left y-axis, black) and corresponding coefficient of variation (CV; right y-axis, red) of Dynabeads in RPMI medium (open squares) and Jurkat cell suspension (closed squares). Error bars represent the standard deviation of the mean of six replicates. Coefficient of variation values ​​(%) were calculated from six replicates. [Figure 10A] The effect of Dynabeads on the total number of cells quantified is shown. [Figure 10B] The effect of Dynabeads on the total number of cells quantified is shown. [Figure 10C]The effect of Dynabeads on the number of total cells quantified is shown. Figure 10A shows cell concentration determined by using CNN (black bars) and particle morphological parameters (empty bars) in samples with different concentrations of added Dynabeads. Figure 10B shows debris concentration determined by using CNN (black bars) and particle morphological parameters (empty bars) in samples with different concentrations of added Dynabeads. Figure 10C shows cell adducts determined by using CNN (black bars) and particle morphological parameters (empty bars) in samples with different concentrations of added Dynabeads. Error bars represent the standard deviation of the mean concentration from triplicate measurements within a day. [Figure 11A] FIG. 13 shows the error rate of particles classified as cells or adducts in Dynabeads suspensions with different Dynabeads target concentrations as determined using FlowCam. [Figure 11B] Figure 1 shows the error rate of Dynabeads and adducts in a cell suspension of 375000 cells / mL as determined using FlowCam. The error rate is based on misclassified particles by using CNN (black bars) and misclassified and unclassified particles by using particle morphological parameters (white bars). Error bars are standard deviation of the mean of sextuplicates. [Figure 12] Figure 1 shows the probability distributions (binned in units of 0.01) of classified particle images determined by FlowCam-CNN for a Dynabeads suspension at a concentration of 80,000 beads / mL (filled) and a cell suspension supplemented with Dynabeads (empty). Representative images are shown with their assigned probability of belonging to a Dynabeads class. [Figure 13A] FIG. 1 shows the concentration of Dynabeads in RPMI medium (control; filled bars) and in cell suspension (empty bars). [Figure 13B] FIG. 1 shows the concentration of Dynabeads in RPMI medium (control; filled bars) and in cell suspension (empty bars). [Figure 13C]Figure 13 shows the concentration of Dynabeads in RPMI medium (control; filled bars) and cell suspension (empty bars). Three different cell concentrations (x-axis; determined using hemocytometry) were tested, with reference concentrations of 20,000 (Figure 13A), 50,000 (Figure 13B) and 100,000 (Figure 13C) Dynabeads / mL. Error bars represent the standard deviation of triplicate measurements of Dynabeads in cell suspension and non-multiple measurements of Dynabeads in RPMI medium. [Figure 14A] FIG. 13. Manual gating of freshly thawed cells. [Figure 14B] FIG. 13. Manual gating of heat-treated cells. [Figure 14C] FIG. 13. Manual gating of camptothecin-treated cells by using Annexin V-FITC and propidium iodide (PI) assays. [Figure 14D] FIG. 13. Manual gating of freshly thawed cells. [Figure 14E] FIG. 13. Manual gating of digitonin-treated cells measured by using the calcein-AM assay. [Figure 15A] FIG. 1 illustrates example FlowCam images selected for training a CNN model. [Figure 15B] FIG. 11 shows a confusion matrix obtained from a cross-validation analysis on an image dataset that was not used during model training. [Figure 15C] FIG. 14. Particle size distribution of vortexed (enriched in debris particles), heat-treated (enriched in necrotic cells) and freshly thawed (viable cells) cell suspensions measured using FlowCam. [Figure 16A] FIG. 1 shows measurements of cell aliquots after thawing at 5° C., 25° C. and 37° C. [Figure 16B] FIG. 1 shows measurements of cell aliquots after thawing at 5° C., 25° C. and 37° C. [Figure 16C] FIG. 1 shows measurements of cell aliquots after thawing at 5° C., 25° C. and 37° C. [Figure 16D]Figure 16 shows measurements of cell aliquots after thawing at 5°C, 25°C and 37°C. Figure 16A shows the determination of the concentration of cell debris, live cells and necrotic cells measured by using FIM-CNN (FIM) and automated cell counting (ACC). Figure 16B shows the measurement of cell viability by performing calcein-AM (CA), propidium iodide (PI) and FIM-CNN assays. Figure 16C shows the assessment of cell apoptosis by Annexin V-FITC assay. Figure 16D shows the assessment of cell apoptosis by LDH release. Error bars represent the standard deviation of the mean. [Figure 17A] FIG. 1 shows the effect of thawing on the size distribution of Jurkat cells. [Figure 17B] FIG. 17 shows the effect of thawing on the size distribution of Jurkat cells. FIG. 17A shows cell diameters measured using FIM after thawing at 5° C. and 37° C. FIG. 17B shows an exemplary image of particles detected in a Jurkat cell suspension after thawing at 5° C. [Figure 18A] FIG. 1 shows cell suspensions supplemented with three different concentrations of DMSO either before one freeze-thaw step (T0), after 3 h of storage at −18° C. (T1) or after 3 h of storage at RT (control). [Figure 18B] FIG. 1 shows cell suspensions supplemented with three different concentrations of DMSO either before one freeze-thaw step (T0), after 3 h of storage at −18° C. (T1) or after 3 h of storage at RT (control). [Figure 18C] Figure 18 shows cell suspensions supplemented with three different concentrations of DMSO either before (T0), after 3 h storage at -18°C (T1) or after 3 h storage at RT (control). Figure 18A shows the concentration of total (viable + necrotic + apoptotic) cells measured by FIM. Figure 18B shows the concentration of debris particles. Figure 18C shows cell viability determined by using calcein-AM (CA), propidium iodide (PI) and FIM-CNN (FIM) assays. Error bars represent standard deviation of the mean. [Figure 19A]FIG. 1 shows cell suspensions supplemented with three different concentrations of dimethyl sulfoxide (DMSO) either before one freeze-thaw step (T0), after 3 h of storage at −18° C. (T1) or after 3 h of storage at RT (control). [Figure 19B] Figure 19 shows cell suspensions supplemented with three different concentrations of dimethyl sulfoxide (DMSO) either before one freeze-thaw step (T0), after 3 h storage at -18°C (T1) or after 3 h storage at RT (control). Figure 19A shows the LDH release assay. Figure 19B shows the apoptotic cell fraction quantified by using imaging flow cytometry (Annexin V-FITC assay). Error bars represent the standard deviation of the mean. [Figure 20A] FIG. 2 shows cell suspensions without or with FBS (10% (v / v)) before (T0) and after shaking (T1) or resting storage (control) at room temperature. [Figure 20B] FIG. 2 shows cell suspensions without or with FBS (10% (v / v)) before (T0) and after shaking (T1) or resting storage (control) at room temperature. [Figure 20C] Figure 20 shows cell suspensions without or with FBS (10% (v / v)) before (T0) and after (T1) shaking or resting storage at room temperature (control). Figure 20A shows the total concentration of cells (viable and necrotic) and debris particles by using FIM. Figures 20B and 20C show LDH release (Figure 20B) and apoptotic cell fraction (Figure 20C) quantified by using imaging flow cytometry (Annexin V-FITC assay) at given time points. Error bars represent standard deviation of the mean. [Figure 21] FIG. 1 shows cell viability measurements by calcein-AM (CA), propidium iodide (PI) and FIM-CNN (FIM) assays after horizontal shaking in the absence and presence of 10% (v / v) FBS. Error bars represent standard deviation of the mean. [Figure 22]FIG. 1 shows the confusion matrix of a CNN retrained to classify live cells, necrotic cells, apoptotic cells and debris particles, where the apoptotic cells used for retraining were induced by using camptothecin. [Figure 23] FIG. 1 shows the percentage of apoptotic cells measured by using a fluorescence-based flow cytometry assay and FIM-CNN in cell suspensions with and without camptothecin (incubation of cell suspensions at 37° C. and 5% CO2) over a 14 hour time course. [Figure 24] FIG. 1 shows images from four categories: viable (healthy) cells, necrotic cells, apoptotic cells and debris. [Diagram 25] FIG. 1 shows a confusion matrix of four categories: viable (healthy) cells, necrotic cells, apoptotic cells and debris. [Figure 26] Figure 1 shows that CNNs underestimated the number of live cells compared to a fluorescence-based flow cytometry assay. Most likely, CNNs misclassified healthy cells as apoptotic due to "overselection" of live cells that were in focus, while live cells that were out of focus were omitted from the training dataset. [Figure 27] FIG. 13 CNN estimation of cell number after removing out-of-focus images from the test dataset. [Figure 28A] FIG. 14 Classification of live and necrotic cells by a CNN trained on brightfield images and calcein-AM staining (fluorescence criteria) obtained by imaging flow cytometry. [Figure 28B] Figure 28 shows classification of live and necrotic cells by CNN trained on brightfield images and calcein-AM staining (fluorescence criteria) obtained by imaging flow cytometry. Figure 28A shows a sample treated with digitonin to induce necrosis. Figure 28B shows freshly thawed cells. [Figure 29A]FIG. 14. Classification of apoptotic and live cells by a CNN trained on brightfield images and Annexin V-FITC / PI staining (fluorescence criteria) obtained by imaging flow cytometry. [Figure 29B] Figure 29 shows classification of apoptotic and live cells by a CNN trained on brightfield images and Annexin V-FITC / PI staining (fluorescence criteria) obtained by imaging flow cytometry. Figure 29A shows a sample treated with camptothecin to induce apoptosis. Figure 29B shows an untreated sample as a negative control. [Diagram 30] FIG. 1 shows the confusion matrix of a CNN classifying viable (healthy), necrotic and apoptotic cells. [Diagram 31] Figure 23 shows the confusion matrix of a CNN classifying viable (healthy), necrotic, apoptotic, and debris. The CNN was trained on images from the optimized training image dataset of the preparation containing viable, apoptotic, and debris used in Figure 25, and a fresh preparation of necrotic cells. [Diagram 32] FIG. 1 shows the distribution of predicted classes obtained by a CNN classifier trained on two different training datasets (1-01 and 1-02) of necrotic cells, where the distribution of classes determined by fluorescence-based flow cytometry is used as reference. [Diagram 33] FIG. 1 shows a confusion matrix of CNN-classified live, necrotic, apoptotic cells and debris incubated in isotonic (Iso), hypotonic (Hypo) and hypertonic (Hyper) conditions. [Diagram 34] Figure 1 shows a comparison of fluorescence-based flow cytometry assays with predictions of CNN-classified live cells incubated in isotonic solution (2-00). Live cells represent the sum of Iso, Hypo and Hyper classes. [Figure 35A] FIG. 13 shows CNN classification of samples with cells incubated for 5 min prior to FlowCam measurement at osmolalities ranging from 220 to 450 mOsm / kg. [Figure 35B]Figure 35 shows CNN classification of samples with cells incubated for 5 min prior to FlowCam measurement at osmolalities ranging from 220 to 450 mOsm / kg. Figure 35A shows classification into Hyper, Iso, Hypo, Necrotic and Apoptotic classes. Figure 35B shows classification of live cells. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0021] The present subject matter may be more readily understood by reference to the following detailed description, which forms a part of this disclosure: It is to be understood that the present disclosure is not limited to the specific products, methods, conditions or parameters described and / or illustrated herein, and that the terminology used herein is for the purpose of describing particular embodiments by way of example only, and is not intended to limit the disclosure as claimed.

[0022] Unless otherwise defined herein, scientific and technical terms used in connection with this application shall have the meanings commonly understood by those of ordinary skill in the art. Further, unless otherwise required by context, singular terms shall include the plural and plural terms shall include the singular.

[0023] In this disclosure, the singular forms "a," "an," and "the" include plural references, and reference to a particular numerical value includes at least that particular value unless the context clearly dictates otherwise. As used herein, the term "plural" means two or more. When values ​​are expressed as approximations, using the antecedent "about," it is understood that the particular value forms another embodiment. All ranges are inclusive and combinable. In some embodiments, the term "about" refers to a deviation between 0.0001 and 10% from the stated number or range of numbers. In some embodiments, the term "about" refers to a deviation of up to 25% from the stated number or range of numbers. The term "comprises" means to include all recited elements, but may also include additional unspecified elements, and may be used interchangeably with the terms "encompasses," "includes," or "contains," all of the same quality and meaning. The term "consisting of" means to consist of the recited elements or steps, and may be used interchangeably with the term "consisting of," all of the same quality and meaning.

[0024] The present disclosure relates to a method for quantifying single viable cells and / or particulate impurities in a cell-based product sample, the method comprising: a. providing a training set of classified flow imaging microscopy (FIM) images including single live cells, debris, and at least one of necrotic cells, apoptotic cells, doublet cells, magnetic beads, or adducts; b. training a convolutional neural network (CNN) using the classified images of step (a); c. Capturing a FIM image from the product sample; d. Identifying single live cells, debris, and at least one of necrotic cells, apoptotic cells, doublet cells, magnetic beads, or adducts by applying the trained CNN obtained in step (b) to the FIM image of step (c); e. Quantifying the single viable cells and / or particulate impurities identified in step (d); Includes.

[0025] Cellular and Particulate Impurities In some embodiments, the cells include any type of cell used in medicine. In some embodiments, the cells include any type of cell used in cancer immunotherapy. In some embodiments, the cells include natural killer (NK) cells, lymphokine-activated killer cells, T cells, cytotoxic T cells, memory T cells, or dendritic cells. In some embodiments, the cells include genetically engineered cells. In some embodiments, the cells include cells expressing chimeric antigen receptors (CARs). In some embodiments, the cells include CAR-T cells.

[0026] In some embodiments, the cells include any type of cell used in regenerative medicine. In some embodiments, the cells include stem cells, progenitor cells, differentiated cells, mesenchymal stem cells, hematopoietic stem cells, induced pluripotent somatic cells (iPSCs), adipocytes, cardiomyocytes, chondrocytes, pancreatic islet cells, or insulin-producing cells. In some embodiments, the cells include cells used in gene therapy. In some embodiments, the cells include somatic cells.

[0027] In some embodiments, the cells comprise allogeneic or autologous cells, hi some embodiments, the cells comprise human or animal cells.

[0028] Those skilled in the art will understand that CARs include engineered receptor proteins that give T cells the ability to target specific proteins. These receptors combine both antigen binding and T cell activation functions into a single receptor. By expressing CARs in this way, T cells can bind to the antigen recognized by the CAR and become activated in its presence.

[0029] Those skilled in the art will understand that a viable cell refers to a cell that can grow in culture or in any suitable medium. This is in contrast to a non-viable cell, which is a dead cell that cannot grow and / or reproduce. Many methods are available in the art for determining whether a cell is viable, including colorimetric methods, Coulter counters, flow cytometry, and morphology-based methods.

[0030] In some embodiments, the live cell images used in the training set include images of live cells grown under similar culture conditions. Experimental data provided herein shows that the accuracy of the CNN classifier can be increased by training the CNN with heterogeneous images within each category. Thus, in some embodiments, the live cell images in the training set include images of live cells suspended in solutions with different osmolalities. In some embodiments, the image training set includes images of live cells suspended in isotonic conditions, live cells suspended in hypotonic conditions, and live cells suspended in hypertonic conditions. In some embodiments, the CNN classifies all live cells suspended under different osmolalities into a single category of live cells.

[0031] In some embodiments, the hypotonic solution contains a lower concentration of solutes than cells. In some embodiments, the hypotonic solution contains an osmolality of 220, 250, or 270 mOsm / kg. The hypotonic osmolality can be obtained, for example, by diluting phosphate buffered saline (PBS) in water. In some embodiments, the hypertonic solution contains a higher concentration of solutes than cells. In some embodiments, the hypertonic solution contains an osmolality of 350, 400, 420, or 450 mOsm / kg. The hypertonic osmolality can be obtained by adding different amounts of sucrose to PBS. In some embodiments, any isotonic cell medium, for example PBS, can be used to incubate cells in isotonic conditions.

[0032] One of skill in the art will appreciate that the term particulate impurities may refer to any undesirable molecules present in a cellular product. In some embodiments, particulate impurities include non-cellular impurities. In some embodiments, non-cellular impurities include magnetic beads.

[0033] One of skill in the art will appreciate that several magnetic beads are available in the art for magnetically sorting cells. Any of these can be used in the methods disclosed herein. In some embodiments, the magnetic beads include Dynabeads. In some embodiments, the magnetic beads include magnetic microparticles coated with antibodies. In some embodiments, the magnetic beads include magnetic nanoparticles coated with antibodies.

[0034] In some embodiments, the particulate impurities include cellular impurities. In some embodiments, the cellular impurities include debris. In some embodiments, the cellular impurities include necrotic cells. In some embodiments, the cellular impurities include apoptotic cells. In some embodiments, the cellular impurities include doublet cells.

[0035] As used herein, the term "doublet cells" refers to two cells captured in a single microscopic image. In some embodiments, doublet cells refer to two cells that are physically bound together. In some embodiments, doublet cells refer to two cells in close proximity that are captured in a single image and detected as if they were a single cell. If doublet cells are not physically bound together, they are not necessarily considered pharmaceutical or clinical impurities. However, unbound doublet cells are considered impurities herein because they prevent accurate quantification of the cells. In some embodiments, identified doublet cells are counted as two cells for the purpose of reporting the determined total concentration of cells.

[0036] The methods disclosed herein can be applied to quantify cells and impurities in any composition that includes cells. In some embodiments, the methods are applied to quantify cells and impurities in a cell-based product sample. Those skilled in the art will understand that cell-based products include pharmaceutical products, drugs or medicines, i.e., compositions used to treat, cure, prevent or diagnose disease or promote well-being. Additionally, cell-based products include approved drugs, drug candidates, or products that are being researched, tested, stored, developed, labeled, manufactured, marketed, sold and / or distributed by pharmaceutical companies.

[0037] One of skill in the art will appreciate that a sample may include any portion of the cell-based product that is analyzed. In some embodiments, a sample includes a volume of a cell product obtained randomly to represent the composition of the cell-based product. In some embodiments, a sample includes multiple samples taken at different time points or from different portions of the cell product. In some embodiments, a sample includes the entire cell product.

[0038] Convolutional Neural Networks The methods disclosed herein include the use of neural networks to identify cells and / or impurities in cell-based products. Those skilled in the art will understand that a neural network or artificial neural network refers to a collection of connected nodes called neurons interconnected by edges. A description of neural networks and their use for image recognition can be found, for example, in Valueva MV et al. (2020). Application of the residue number system to reduce hardware costs of the convolutional neural network implementation. Mathematics and Computers in Simulation. Elsevier BV. 177: 232-243. doi: 10.1016 / j.matcom.2020.04.031. ISSN 0378-4754. The neurons and edges have weights that are adjusted during a learning or training process. In the present disclosure, the learning includes providing a set of classified images of single live cells, debris, necrotic cells, apoptotic cells, doublet cells, magnetic beads, and / or adducts. Typically, the neurons are aggregated into layers. A signal, as a microscope image, travels from the first layer (input layer) to the final layer (output layer), possibly after traversing the layers multiple times.

[0039] A deep neural network is a neural network that has multiple layers between an input layer and an output layer. In some embodiments, the methods disclosed herein include the use of a deep neural network. A convolutional neural network (CNN) is a type of deep neural network that is commonly applied to the analysis of visual images. A CNN is a multi-layer fully connected network, i.e., each neuron in one layer is connected to every neuron in the next layer.

[0040] In the methods disclosed herein, the input layer of the CNN is fed with a magnified image of a cell-based product sample. In some embodiments, the image is obtained by flow imaging microscopy (FIM). One skilled in the art will appreciate that FIM, sometimes referred to as "dynamic image analysis," includes any instrument consisting of a microscope arranged and configured to capture an image of a flow cell. Numerous FIM systems are available in the art, any of which can be used with the methods disclosed herein. In some embodiments, the FIM includes FlowCam® (Yokogawa Fluid Imaging Technologies Inc.). In some embodiments, the FIM is selected from the group including MFI™ (ProteinSimple), iSpect IDA-10 (Shimadzu), Occhio (Occhio Instruments SA), and FlowSight® Imaging Flow Cytometer (Merck).

[0041] In some embodiments, obtaining the flow imaging microscopy image includes capturing an image of the field of view with a digital camera and optionally segmenting the image with a grayscale thresholding process to create a binary image of each particle, hi some embodiments, the image is obtained by a light microscope, a fluorescent microscope, a polarizing microscope, a phase contrast microscope, or a differential interference contrast microscope.

[0042] In some embodiments, the methods disclosed herein quantify the number of single live cells, necrotic cells, apoptotic cells, doublet cells, magnetic beads, adducts, or debris in an analyzed sample. In some embodiments, the quantification comprises assigning each image to a category, e.g., single live cells, apoptotic cells, etc., and adding the number of images assigned to each category. In some embodiments, the quantification comprises providing the number of single live cells and / or impurities in the sample. In some embodiments, the quantification comprises calculating a probability function for each of the analyzed images. In some embodiments, the quantification comprises adding the probability functions of all the analyzed images.

[0043] Those skilled in the art will understand that CNNs for image classification can include different architectures. Any of these architectures can be used in the methods disclosed herein. In some embodiments, the CNN includes the VGG-19 architecture, a description of which can be found, for example, in Simonyan K et al. Very deep convolutional networks for large-scale image recognition; 2014. http: / / arxiv.org / pdf / 1409.1556v6. VGG-19 is a neural network architecture that includes a depth of 19 layers and an image input size of 224x224. The VGG-19 CNN was pre-trained on over 1 million images from the ImageNet database. The VGG-19 CNN has been shown to classify images into 1000 object categories.

[0044] In some embodiments, only a portion of the CNN connection layers are active and the rest are frozen. Any number of layers can be active during training with the cell image dataset. In some embodiments, the last two fully connected layers are active during training. In some embodiments, the first 10 layers are frozen and the rest are active.

[0045] Numerous algorithms can be used to train a CNN, any of which can be applied to train the CNN disclosed herein. In some embodiments, the CNN is trained with the Adam optimization algorithm. In some embodiments, the CNN is trained with the SGD optimization algorithm.

[0046] Training Set The methods disclosed herein include training a CNN with a set of classified images from different categories. In some embodiments, the categories are selected from the group including single live cells, necrotic cells, apoptotic cells, doublet cells, magnetic beads, adducts, and debris. In some embodiments, the training set includes images from 2, 3, 4, 5, 6, 7, or more categories.

[0047] In some embodiments, the training set includes single live cells, necrotic cells, apoptotic cells, doublet cells, magnetic beads, adducts, and debris. In some embodiments, the training set includes single live cells, doublet cells, magnetic beads, adducts, and debris. In some embodiments, the training set includes single live cells, necrotic cells, and debris. In some embodiments, the training set includes single live cells, necrotic cells, apoptotic cells, and debris.

[0048] For a CNN to learn to accurately distinguish categories, it must first be trained with a large number of images that are correctly assigned to that category. Mislabeling of images in the training set will degrade the CNN's future ability to identify members of that category. For example, labeling live cells as apoptotic will impair the CNN's future ability to differentiate between live and apoptotic cells. While the correct labeling of training images may seem trivial, in practice it raises many complex obstacles. A cell culture contains millions of cells, which respond differently to the same external conditions. For example, treatment of a cell population with a pro-apoptotic agent usually induces apoptosis in only a small portion of the population. Labeling all cells that have undergone such a pro-apoptotic treatment as "apoptotic" and using that label to train a CNN would be detrimental to the performance of the CNN. In this case, the ability of the CNN to distinguish between apoptotic and non-apoptotic cells would be reduced. b. In principle, numerous methods are available to confirm whether cells are apoptotic, such as labeling cells with apoptotic markers. Although these methods are widely used in static microscopy, they cannot be applied to FIM due to the inherent properties of the system. In FIM, images are taken as the flowing fluid passes through a flow cell, and fluorescent images (to confirm the nature of the cells) cannot be acquired simultaneously with bright-field images (to train the CNN). c. Thus, the solution proposed herein involves obtaining a highly homogenous population of live, necrotic, apoptotic, or adduct cells, and assigning all of the cells to the corresponding category. The population is homogenous enough that the CNN can tolerate mislabeling and still achieve high classification accuracy.

[0049] The discovery presented herein that it is possible to accurately train a FIM-CNN using the training set disclosed herein is surprising and unexpected.

[0050] Although many different image training sets were tested, we were unable to adequately train the CNN. A suitable image set was obtained only after extensive experimentation and analysis, as detailed in the Examples section.

[0051] In some embodiments, the training set comprises: i) providing a cell suspension; ii) applying said cell suspension conditions to induce necrosis, apoptosis, cell debris, appendages, or any combination thereof in said cell culture; iii) optionally labeling cells and other particulate impurities; iv) acquiring a FIM image of the cell culture; and v) classifying said FIM images; and The method is obtained by a method comprising the steps of:

[0052] In some embodiments, the cell type in step (i) is the same as the cell type of the cell-based culture sample. In some embodiments, the cell suspension in step (i) is the same as the cell suspension of the cell-based product. One skilled in the art will appreciate that the methods disclosed herein are applicable to any type of cell suitable for a cell-based product.

[0053] In some embodiments, the cell culture conditions for inducing necrosis include subjecting the cells to stress conditions. In some embodiments, the stress conditions include shaking stress. In some embodiments, the shaking stress includes agitation at about 185 rpm at room temperature for about 3 hours.

[0054] In some embodiments, the stress condition comprises a freeze-thaw cycle, which in some embodiments comprises adding a solvent, such as DMSO, to the cell culture sample at concentrations of 1%, 2.5%, 5% and 10%, then freezing the sample at -18°C for at least about 3 hours, and then thawing the sample at 37°C for 2 minutes.

[0055] In some embodiments, the stress conditions include exposing the cell culture to heat stress. In some embodiments, the heat stress includes exposing the cells to about 55° C. for about 90 minutes.

[0056] In some embodiments, the stress condition comprises incubation with ethanol, hi some embodiments, incubation with ethanol comprises adding ethanol to the cell culture sample to a concentration of 10% for 90 minutes at 37°C.

[0057] In some embodiments, the stress conditions include a combination of two or more of shaking stress, freeze-thaw cycles, heat and ethanol incubation.

[0058] In some embodiments, the CNN is trained on images of necrotic cells obtained by different methods. For example, the CNN can be trained on images of necrotic cells obtained by incubating a first group of cells in 10% ethanol at 37° C. for 90 minutes, and also on images of necrotic cells obtained by incubating a second group of cells at 55° C. for 90 minutes.

[0059] In some embodiments, the cell culture conditions for inducing apoptosis include incubating the cells with a pro-apoptotic reagent. In some embodiments, the pro-apoptotic agent includes about 1 μM staurosporine, where the cells may be incubated for about 3 hours.

[0060] In some embodiments, the cell culture conditions for inducing adducts include incubating cells in the presence of Dynabeads at a population ratio of about 1:1 for about 1.5 hours.

[0061] In some embodiments, the cells used in the training set images are labeled. One skilled in the art will appreciate that numerous markers are available for labeling the cellular phenotypes and / or impurities investigated herein. In some embodiments, the assay for determining apoptosis includes labeling cells with Annexin V, DNA stain, activated caspase detector, cytoplasmic cytochrome C, glutathione, or combinations thereof. In some embodiments, the assay for determining necrosis includes labeling cells with, for example, RIPK3 or MLKL. In some embodiments, morphological filters such as intensity, sigma intensity, convexity, compactness, and aspect ratio can be used to identify live cells, necrotic cells, apoptotic cells, debris, beads, and adducts.

[0062] In some embodiments, the training set images are obtained by flow imaging microscopy (FIM). Those skilled in the art will appreciate that FIM is a system that allows for obtaining microscopic images of a liquid sample in a flow. FIM uses an ultra-precise computer-controlled syringe pump to draw a fluid sample through a flow cell perpendicular to the optical path. An optical system similar to a microscope is used to capture real-time images of particles in a fluid as they pass through the flow cell.

[0063] In some embodiments, the cell images of the training set are classified into categories according to the treatment applied to the cells. For example, all cell images from cell cultures treated with a pro-apoptotic agent are classified as apoptotic cells. In some embodiments, the cell images are classified manually. In some embodiments, the cells are classified into categories according to predetermined morphological parameters. In some embodiments, the morphological parameters are measured by associated software for analyzing the microscopic images. In some embodiments, at least one parameter selected from intensity, sigma intensity, convexity, compactness, and aspect ratio is used to assign the cell images to categories. In some embodiments, each cell image is visually analyzed and manually assigned to a category.

[0064] In some embodiments, images of live cells are classified by applying a size filter (10 μm-20 μm) and an aspect ratio filter (0.6-1.0). In some embodiments, the classification of images of live cells is further manually refined by excluding visible apoptotic blebs and cells with rougher surfaces.

[0065] In some embodiments, all images from cells that have been subjected to conditions for apoptosis induction are classified as apoptotic cells. In some embodiments, cell images are classified into the apoptotic cell category by applying a size filter (10-20 μm) that excludes debris and aggregates, and a circularity filter (<0.9) that excludes remaining live cell images.

[0066] In some embodiments, all images from cells that have undergone conditions for necrotic induction are classified as necrotic cells. In some embodiments, manually selected images of live and necrotic cells are filtered according to morphological parameters such as aspect ratio, circular fit, convexity, sigma strength and symmetry. For each listed morphological parameter, a particle (cell) must have a value between the 10th and 90th percentile of the manually selected population (live or necrotic cell population) to be selected for input into training the model.

[0067] In some embodiments, all images from cells that have been subjected to conditions to form adducts are classified as adduct cells. In some embodiments, images from adducts are classified based on average particle intensity values. In some embodiments, separation of single cells from doublet cells is performed based on an aspect ratio parameter.

[0068] In some embodiments, images from magnetic beads and adducts are sorted according to their low transparency value. In some embodiments, images of cells or debris are selected according to their high transparency value. In some embodiments, image selection is manually double-checked to exclude debris particles.

[0069] In some embodiments, the training set classification images are obtained by a method that essentially includes obtaining a sample of a cell-based product, obtaining a FIM image of the sample, and classifying the FIM image by any of the classification methods disclosed herein.

[0070] Those skilled in the art will appreciate that neural networks can be pre-trained on common image datasets. In some embodiments, pre-training a CNN involves a process in which the weights of the network are adjusted to provide a good starting point for training the CNN with a training set. In some embodiments, the CNN is pre-trained on the ImageNet dataset, which can be downloaded from http: / / image-net.org.

[0071] Image Classification Following learning of the categories of interest, the CNN can classify new images into one of those categories. In some embodiments, the input layer of the CNN is fed with images from a cell-based product sample, and the images are classified into a category selected from single live cells, debris, necrotic cells, apoptotic cells, doublet cells, magnetic beads, or adducts. In some embodiments, the CNN output includes a probability function, the probability function including a probability of the analyzed image belonging to each category. In some embodiments, the CNN output includes an assignment of the analyzed image to a category.

[0072] In some embodiments, the methods disclosed herein include applying a function to the results provided by the CNN. In some embodiments, the function includes predicting the number of viable cells and / or impurities in the product-cell sample. In some embodiments, the methods disclosed herein are part of a quality control process and can be combined with any other quality control analysis known in the art.

[0073] In some embodiments, the methods disclosed herein include labeling the images of the training set with a fluorescent marker. One of skill in the art will appreciate that cell markers are available in the art for fluorescently labeling live, necrotic and apoptotic cells. Any of these markers can be implemented in the methods disclosed herein.

[0074] In some embodiments, the FIM simultaneously acquires brightfield images and fluorescent signals of the cell suspension. The fluorescent signals are then used to label the brightfield images. In some embodiments, the fluorescent signals indicate whether the cells are viable, apoptotic, or necrotic, and the brightfield images are classified accordingly and labeled accordingly. In some embodiments, the labeled brightfield images are used to train a CNN.

[0075] Those skilled in the art will understand that any cell marker that indicates any cell phenotype can be applied to the methods disclosed herein. Thus, for example, the methods disclosed herein can be used to train CNN to recognize cells that express any membrane protein. EXAMPLES

[0076] The following examples are presented to more fully illustrate some embodiments of the technology disclosed herein, but they should not be construed as limiting the scope of the invention in any way.

[0077] Materials and Methods for Examples 1 to 12 Unless otherwise stated, the materials and methods described below relate to Examples 1-12.

[0078] Cell and sample preparation. T-cell leukemia cells (Jurkat, clone E6-1, ATCC® TIB152™) were obtained from Leiden University Medical Centre (LUMC) at a total cell concentration of 10 7Cells were provided as frozen 1 mL aliquots of cells / mL and stored at -140°C in a freezer prior to use. Jurkat cells were formulated in high glucose RPMI 1640 (RPMI medium; ThermoFisher, Waltham, MA, USA) supplemented with 10% fetal bovine serum (FBS; Life Technologies, Eugene, OR, USA) and 10% dimethyl sulfoxide (DMSO) (Life Technologies). Dynabeads Human T-Activator CD3 / CD28 for T Cell Expansion and Activation, and low protein binding collection tubes were purchased from ThermoFisher. Sterile 5 mL Eppendorf tubes were purchased from VWR (Ismaning, Germany).

[0079] Jurkat cells used as model T cells in this study were thawed and freshly prepared in RPMI medium prior to analysis. Frozen cell aliquots were thawed at 36°C and resuspended in approximately 40 mL of RPMI medium. To remove residual FBS and DMSO, the cell suspension was centrifuged at 300 rcf for 10 min at 20°C. Unless otherwise stated, the supernatant was removed and the pellet was resuspended in 10 mL of RPMI medium. Unless otherwise stated, the mean concentration of cells (live and dead) was 477,188 ± 85,914 cells / mL, and the mean viability, determined by hemocytometry (described below), was 81% ± 9% (n = 8). Cell-containing samples were measured up to 4 hours after thawing, during which time cell viability was unaffected.

[0080] Dynabeads were incubated for 10 min in RPMI medium. 6The intermediate stock was diluted to a concentration of beads / mL (based on the dilution factor of the nominal Dynabeads concentration) and stored at 2-8 °C for up to 1 month. The required amount of intermediate stock was added to the cell sample to achieve the desired Dynabeads concentration. The reference concentrations of Dynabeads described in the examples below are the expected concentrations of Dynabeads in the samples obtained from dilution calculations and the original bead concentrations stated by the manufacturer. It must be noted that the manufacturer does not use a FlowCam for quantification of Dynabeads, therefore systematic deviations between the reference and measured concentrations should be expected.

[0081] Hemocytometry Cell viability and total cell concentration were determined by using Bright-Line hemocytometer glasses (Merck, Darmstadt, Germany) and an Axiostar Plus microscope (Zeiss, Jena, Germany) at 10x magnification (Zeiss). The washed cell suspension was diluted 2-fold with sterile filtered 0.4% trypan blue solution (Merck, Darmstadt, Germany). Then, 10 mL of the mixture was placed in a hemocytometer and at least 100 cells were counted (both viable and non-viable stained cells) according to the manufacturer's recommendations.

[0082] Flow Imaging Microscopy For the characterization of micrometer-sized particles, a FlowCam 8100 (Fluid Imaging Technologies, Scarborough, ME, USA) equipped with an 80 mm flow cell and a 10x objective was used. The instrument was operated by using VisualSpreadsheet software (v4.10.8). Analysis was performed at a flow rate of 0.18 mL / min with a detection threshold of 17 dark pixels and 15 light pixels. Images were captured at 27 frames / s using a high-resolution CMOS camera (1920x1200 pixels). A total sample volume of 0.5 mL was analyzed with an efficiency of approximately 70% (i.e., the measured sample volume was −0.35 mL). The cleaning steps between sample measurements included thoroughly washing the flow cell with 2% Hellmanex III and highly purified water. Diameters are reported as equivalent sphere diameters (ESDs) and no filters were applied for imaging pretreatment. Samples were measured in triplicate or sextuplicate unless otherwise stated.

[0083] The samples measured within this study contained particles of five different populations: single cells, doublet cells, Dynabeads, adducts (defined as a combination of at least one bead and at least one cell) and debris (particles of any other cell type and non-cell type). To report the total determined concentration of cells, the counts of single cells, adducts and 2x doublet cells were summed. The determined concentration of Dynabeads in the measured samples was derived from the total number of Dynabeads and adducts.

[0084] Generation of particle images for population discrimination Establishment of thresholds and training of the CNN were performed on manually selected images (4,000-4,500) of each population class, which was shown to be sufficient to train the model to reach a high classification accuracy (>0.99). To facilitate the selection process of debris and adducts, samples with an increased number of the respective particles were generated prior to FlowCam analysis. Samples enriched in particles representing debris were obtained by subjecting freshly resuspended (without cryoprotectant) cell suspensions to two freeze-thaw cycles (-140 to 36 °C). Samples with a large number of adducts were created by incubating cells (500,000 cells / mL) in the presence of Dynabeads in a 1:1 number ratio for 1.5 h at 37 °C and 5% CO2.

[0085] Development of morphological filters for FlowCam The VisualSpreadsheet software of the FlowCam system outputs 30 morphological parameters for each detected particle in the measured sample. Five of these available parameters (intensity, sigma intensity, convexity, compactness and aspect ratio) were found to have the highest resolution for particle populations. The values ​​of particle properties for each population class were further clustered into 1 mm size bins and shown in box plots (Figure 1). To develop the thresholds used to assign each particle to its class, an approach such as that reported in Weinbuch D et al.Micro-flow imaging and resonant mass measurement(Archimedes)-complementary methods to quantitatively differentiate protein particles and silicone oil droplets.J Pharm Sci.2013;102(7):2152-65.(https: / / doi.org / 10.1002 / jps.23552) was used for the discrimination of silicone oil and protein aggregate particles. Briefly, a stepwise approach was followed.

[0086] First, Dynabeads and adducts were separated from cells (singles and doublets) and debris based on the mean particle intensity values. For each size bin of 1 μm, the mean values ​​of the 10th quartile of "low transparent" particles (Dynabeads and adducts) and the 90th quartile of "high transparent" particles (cells and debris) were calculated as a function of size. For size regions where only one population was present, the cutoff threshold was manually adjusted below or above the 95th quartile parameter value. Furthermore, a 4th degree polynomial function was fitted to these points between 3 μm and 35 μm, and test particles of a particular diameter above or below the threshold set were assigned to either group. Second, the compactness, convexity and sigma intensity thresholds that allow the separation of cells from adducts and debris from Dynabeads were derived in a similar manner. For the separation of single and doublet cells, the aspect ratio parameter was applied. Thus, each test particle must meet at least three criteria to be assigned to a particular population class. All particles with a diameter less than 3 μm were assigned to the debris population.

[0087] Deep Convolutional Neural Networks The VGG-19 architecture is used as the basis for the CNN, as described in Simonyan K et al.Very deep convolutional networks for large-scale image recognition;2014.http: / / arxiv.org / pdf / 1409.1556v6. This architecture contains 19 convolutional (weight) layers and can capture a wide range of visual object features. The network weights are optimized by reformulating the convolutional layers as learned residual functions, with the input to the layers as the reference. By fine-tuning only the last two fully connected layers, the feature complexity of the pre-trained model can be optimized for the particle classification task. For fine-tuning, the image dataset was split into a training set, a validation set and a test set in the ratios of 0.8, 0.1 and 0.1, respectively. Such a split of the dataset aims to maintain the balance of the classes, so that the fine-tuning is not biased towards a particular class. The deep learning model was fine-tuned for 30 epochs using the Adam optimization algorithm. The machine learning model was implemented in the Keras (2.2.4)-Tensorflow (1.13.1) Python (3.7.3) library and run on an Nvidia Turing GPU with 11 GB of VRAM. The simplified workflow for image analysis by using CNN is shown in Figure 2.

[0088] Data analysis Statistical analysis of the data was performed with Origin 2016 (Origin-Lab Corporation, Northampton, MA, USA). Box plots represent the distribution of data with the central rectangle spanning the first to third quartiles and whiskers ranging from the 5th to the 95th percentile. For comparison of mean values, a two-tailed Student's t-test with a = 0.05 (95% confidence interval) was used. The limits of detection (LOD) and limits of quantification (LOQ) were determined by using the full range of values ​​tested for Dynabeads, with six measurement replicates performed for each bead concentration. LOD and LOQ were calculated using Equations 1 and 2.

[0089] (1) LOD = (3:3 * σ) / S

[0090] (2) LOQ = (10 * σ) / S

[0091] where σ is the standard error of the y-intercept and S is the slope of the linear regression line.

[0092] Example 2 - Protocol for obtaining high recovery of Dynabeads The goal of these experiments was to determine the best protocol for obtaining Dynabeads images for training a CNN.

[0093] Methods: We compared the effectiveness of three setups:

[0094] Setup 1 Homogenization of Dynabeads in the primary package was performed according to the manufacturer's instructions. Dilutions and homogenization of the beads were performed in polypropylene 2.0 mL microcentrifuge tubes using regular Eppendorf pipette tips. Homogenization of each dilution was performed by vortexing.

[0095] Setup 2. Homogenization of Dynabeads in the primary package was performed according to the manufacturer's instructions. Dilutions and homogenization of the beads were performed in polypropylene 2.0 mL microcentrifuge tubes using regular Eppendorf pipette tips. Homogenization of each dilution was performed on a rotator.

[0096] Setup 3. Homogenization of Dynabeads in the primary package was performed according to the manufacturer's instructions. Dilutions of the beads were performed using low protein binding 2.0 mL microcentrifuge tubes and low retention pipette tips. Homogenization of each dilution was performed by vortexing.

[0097] Results: Setup 3 describes the preparation method of Dynabeads suspension. Dynabeads prepared according to this method could be reproducibly quantified with a recovery rate of over 60% (Figure 3).

[0098] Example 3 - Generation of an annotated image The goal of these experiments was to generate a dataset of adnexal images.

[0099] Methods: Cells were incubated with Dynabeads in a 1:1 ratio for 3 h in a cell incubator (37°C, 5% CO2). Cells were resuspended by pipetting and quantified by the FlowCam method. Images of cells attached to one or more Dynabeads were manually selected by the operator. Adducts generated with different amounts of loaded Dynabeads were quantified by FlowCam-CNN.

[0100] Results The number of adducts detected increased with the number of Dynabeads present in the cell suspension (Figure 4).

[0101] Example 4 - Selecting the appropriate focal point for FlowCam measurements The goal of these experiments was to select a proper focal point for FlowCam measurements. Proper focal point should allow for obtaining images of reproducible quality and achieving consistent performance of the machine learning model.

[0102] Methods Mean intensity and edge gradient values ​​were used to aid in the selection of optimal focus settings for capturing cell images. Once the focus settings were established with the cells, replicate measurements of polystyrene standard particles were performed (n=9) to obtain mean intensity and edge gradient values ​​for standard suitability test criteria. Prior to measurements of the cell suspension, it was necessary to perform measurements of the polystyrene standard particles and adjust the focus settings to the predefined mean intensity and edge gradient values.

[0103] Results Cell images with consistent image quality for classification by using machine learning were collected (Figure 5). A system suitability test was established to test and confirm correct focus settings (based on average intensity and edge gradient values) using polystyrene standard particles.

[0104] Example 5 - Selection of images for the training dataset The aim of these experiments was to select images that contained a high percentage of live cells, necrotic cells and debris particles.

[0105] Methods Establishment of cell treatment methods to induce necrosis. Cells were treated with digitonin (target concentrations in RPMI medium: 1 μM, 5 μM, 10 μM), heat (55°C for 1.5 h) or ethanol (10% v / v, 37°C for 1.5 h).

[0106] Morphological criteria were established to control the training dataset and eliminate "edge cases" for each population. Images taken for training were required to have values ​​of the following morphological parameters within the 10th-90th percentile of the entire population of images selected for each class: aspect ratio, circular fit, compactness, convexity, elongation, intensity, roughness, sigma intensity, symmetry, and transparency.

[0107] Results Classification accuracy using the trained machine learning model reached over 96% (Figures 6A-C).

[0108] Example 6 - Identification of particle populations in cell suspensions Analysis of cell suspensions supplemented with Dynabeads was performed by using the FlowCam and representative examples of generated images of five different particle populations. In addition to single cells, debris and Dynabeads, a notable number of images with two trapped cells (doublet cells) as well as cells with one or more adjacent Dynabeads (adducts) were observed. Particle size distributions of samples containing cells (no beads), Dynabeads (no cells) and a mixture of cells and Dynabeads are shown in Figures 7A-7C. The sample containing cells showed a broad peak between 10 μm and 16 μm, representing Jurkat cells (Figure 7A). In addition, a sharp peak was observed at the lower size limit of detection, assigned to debris. Dynabeads showed a bimodal peak with maxima at 3.5 μm and 6.0 μm (Figure 7B). These values ​​represent the size of the beads measured from focused (sharp) and unfocused (blurred) images and are close to the average bead diameter of 4.5 μm stated by the manufacturer. The particle size distribution of the Dynabeads + cells mixtures appeared to be that of the sum of the cells and beads (Figure 7C). These samples were found to contain appendages (Figure 7, top panel) that were not apparently present in the other samples, but the number of appendages was relatively small and did not substantially affect the overall size distribution.

[0109] Example 7 - Morphological parameters for particle classification For beads suspended in RPMI medium, a linear relationship was found between the measured Dynabeads concentration and the expected Dynabeads concentration derived from the dilution factor, based on selected morphological parameters using FlowCam software, R 2 = 0.95 (Figure 8A). Measurement of beads in the presence of cells (approximately 500,000 total cells / mL using a hemocytometer) yielded a similar linear correlation (R 2= 0.98). However, an unexpectedly large number of unclassified particles (i.e., particles with morphological parameters that did not fall into any of the five classes) were found. The coefficient of variation of the Dynabeads concentrations measured in the presence of cells was significantly higher compared to the control samples. Furthermore, the recovery of Dynabeads in cell suspensions at the lowest three reference concentrations was greater than 100%, suggesting that a large number of debris and other particles were misclassified as Dynabeads (or adducts) when using this approach. Dynabeads suspended in RPMI medium showed recoveries ranging from 25% (lowest Dynabeads concentration) to 85% (highest Dynabeads concentration). Overall, a good correlation was obtained between the detected and reference concentrations of Dynabeads when using the morphological particle parameters output from the FlowCam software. However, the large number of unclassified particles and the significant variation in the determined particle concentrations indicate the fragility of this classification approach.

[0110] Example 8 - CNN for particle classification Conventional morphological filters were unable to accurately identify particle populations in cell suspensions. Therefore, a convolutional neural network (CNN) was applied to the analysis of raw FlowCam images. An average of 140,000 images per sample were classified into individual particle classes using a retrained FlowCam-CNN model with a preselected dataset, as described in Example 1.

[0111] In contrast to the results based on the morphological particle parameter output, the FlowCam-CNN analysis resulted in the classification of all particle-present cell suspensions (500,000 cells / mL counted using a hemocytometer) with or without Dynabeads (Figure 9). Linearity over the entire Dynabeads concentration range tested was greater than 0.95 for both sample sets, with a slope value of 0.8. Samples containing cells and greater than 50,000 beads / mL had recovery values ​​greater than 80% and coefficients of variation less than 15%. The relative errors in the measured bead concentrations were apparently random, showing a uniform distribution around zero for samples containing cells, except for the lowest bead concentration measured. Dynabeads suspended in cell-free RPMI medium at reference concentrations less than 60,000 beads / mL showed lower recoveries compared to samples containing cells, exceeding 75% recoveries within the tested range. The coefficients of variation in the measured Dynabeads concentrations were greater than 10% for samples less than 50,000 beads / mL and less than 10% for the higher tested bead concentrations. Therefore, the optimal Dynabeads concentration for bead quantification by using FlowCam-CNN was determined to be approximately 45,000 beads / mL to at least 200,000 beads / mL.

[0112] The FlowCam method assisted by automated image classification (FlowCam-CNN) was tested in alignment with the ICH Q2(R1) guideline for validation of analytical procedures. The accuracy, precision, LOD, LOQ and linear relationship of Dynabeads detection in the absence and presence of cells were evaluated and the results are shown in Table 1. For Table 1, the LOD and LOQ were determined for the Dynabeads concentrations tested in the study and are shown in Figure 9. The accuracy and precision were determined for Dynabeads concentrations above the LOQ. The cell concentration was approximately 375,000 cells / mL.

[0113] JPEG2024517592000002.jpg54152

[0114] The LOD and LOQ were 15,000 and 45,000 beads / mL, respectively, with slightly higher values ​​found for Dynabeads suspensions in the absence of cells. Accuracy and precision (reproducibility) were calculated for Dynabeads concentrations above the LOQ, and values ​​represent the average of two intraday sets of triplicate measurements. Precision was determined as the percentage recovery spiked with Dynabeads relative to the reference concentration (% recovery). Precision was found to be substantially lower for beads in the absence of cells at reference concentrations below 50,000 beads / mL. Above this concentration, the presence of cells did not affect the accuracy of quantification. Furthermore, the precision of the Dynabeads concentration measurements was evaluated as the coefficient of variation (CV%), with overall values ​​below 10%.

[0115] Example 9 - Characterization of Cellular Particulate Matter In addition to developing methods for the characterization of non-cellular particles, debris (potential impurities) as well as cells and adducts were quantified.

[0116] Test concentrations of Dynabeads in Jurkat cell suspensions (Figure 9) were tested at a constant cell concentration (385,711 ± 59,337 cells / mL as measured by FlowCam). The presence of Dynabeads did not significantly affect the number of total cells quantified (t-test, two-tailed, P>0.17; Figure 10A). Furthermore, the number of detected particles classified as debris in cell samples without and with Dynabeads was highly comparable. Furthermore, the number of detected adducts increased with the higher concentration of beads present in the cell sample (Figures 10B and 10C).

[0117] Example 10 - Misclassification The misclassification rate was calculated indirectly due to the large number of images acquired in one measurement (more than 100,000 in one measurement). Debris particles were present in all measured samples (Dynabeads and cell suspensions) and therefore the misclassification rate of this population was not considered. Figure 11A represents the percentage of incorrectly detected cells (singles, doublets and adducts) and unclassified particles in Dynabeads suspensions at different reference concentrations. Figure 11B shows the error rate of detected Dynabeads and adducts as well as unclassified particles in a cell-only suspension at a cell concentration of 375,000 cells / mL (measured by FlowCam). In both cases, the misclassification rate was very low (less than 2%) when the data were processed using CNN. Moreover, the error rate was independent of the spike-in amount of Dynabeads, since the percentage of incorrectly classified particles was similar for each of the tested concentrations of beads in the cell suspension. Particles analyzed by using morphological parameters showed a much higher imprecision and percentage of unclassified, which is reflected by a relatively high error rate (up to 50 times higher compared to CNN).

[0118] Figure 12 shows the probability distribution determined by the deep learning classification of particle images classified as Dyna beads. Particle images were collected during FlowCam measurements of Dynabeads (80,000 beads / mL) in the presence and absence of cells. In the classification network, we integrated a softmax regression function, which is an activation function that converts the calculated weights into a probability distribution and rejects all cases with a probability less than 0.2. For samples with Dynabeads suspended in RPMI medium, the majority of images classified into the bead class had a probability equal to 1, confirming that the network made the assignment with very high confidence. Moreover, this high confidence was not affected by the presence of cellular material in the sample, since a similar count of beads with a probability of 1 was found in samples containing Jurkat cells. In conclusion, these data demonstrate that image classification by using FlowCam-CNN is highly accurate.

[0119] Example 11 - Effect of cell concentration on measurements The effect of Jurkat cell concentration on the quantification of Dynabeads in cell suspensions by FlowCam-CNN was investigated and the results are shown in Figure 13. Dynabeads added to samples with cell concentrations up to approximately 500,000 cells / mL (as measured using hemocytometry) resulted in similar measured bead concentrations in the presence of cells. At the highest tested cell concentration (900,000 cells / mL), an underestimation of beads detected was observed at reference Dynabeads concentrations of 50,000 and 100,000 beads / mL. Such underestimation was not observed at the lowest tested Dynabeads concentration (20,000 beads / mL). Furthermore, control samples showed lower Dynabeads recovery compared to samples containing cell suspensions.

[0120] Example 12 - Summary of Examples 1-11 Characterization of cellular and non-cellular (i.e., foreign) particulate matter in cell-based medicinal products (CBMPs) is important to ensure a good quality and safe product. Moreover, due to the limited time available for analytical testing of some cell products, simple, rapid and comprehensive methods are urgently needed.

[0121] In this study, we used a model system containing Jurkat cells, which serve as a surrogate for T cells such as CAR-T cells, and Dynabeads CD3 / CD28, which serve as a representative potential process-related particulate impurity in CAR-T cell products. These superparamagnetic beads coupled to CD3 and CD28 monoclonal antibodies are used in the production of CBMPs. However, their removal prior to the final formulation step remains challenging. More importantly, given reports on the potential toxicity of Dynabeads, methods are needed that demonstrate effective and consistent removal of these impurities in the manufacturing process. Therefore, in this study, we evaluated the feasibility of FIM for the identification of Dynabeads and cells and the assessment of Dynabeads concentration. It was hypothesized that this is possible because Dynabeads differ in size and morphological characteristics from T cells.

[0122] FIM-based methods may offer advantages for characterizing CBMPs because they are high-throughput techniques that allow for rapid measurement of a high number of samples without laborious preparation steps. The two most commonly used FIM systems are FlowCam and Micro-Flow Imaging (MFI). A drawback of the FlowCam technique may be the relatively inaccurate particle sizing, resulting from the narrow focal depth in the field of view of the imaging system. As a result, particles of uniform diameter may exhibit a bimodal distribution, which was also observed in this study. Unexpectedly, accurate sizing was not critical for the accuracy of the present system. However, both in-focus and out-of-focus images should be assigned to the same particle class, implying an increased complexity of the classification process.

[0123] Verification of the viability and total concentration of the cells used in the study was performed using a hemocytometer. The cell concentrations determined using FlowCam were lower compared to manual counting, the difference being approximately 20%. Furthermore, the concentrations of Dynabeads detected by FlowCam deviated from the reference concentrations stated by the manufacturer, i.e., the recovery was always less than 100%, especially for the low concentrations of Dynabeads. However, this is most likely due to the loss of beads during sample preparation, since measurements with a Multisizer 4e Coulter-Counter analyzer (method used by the manufacturer for quantification) showed similar results. Furthermore, preliminary studies showed a significant influence of the used laboratory disposables (e.g. low protein binding material, volume to surface ratio) on the determined concentration of Dynabeads. The high affinity of the antibody-coated beads for the surface probably led to the adhesion of Dynabeads to polypropylene tubes and tips used for sample handling. Interestingly, a more consistent and higher recovery was observed for Dynabeads in the presence of cells compared to Dynabeads in cell-free RPMI buffer (control) over the measured concentration range. Notably, a significant loss of beads at low concentrations (below 60000 beads / mL) was observed in cell-free control samples, resulting in less than 50% recovery. Such a high loss of Dynabeads was not observed in cell-containing samples, most likely due to debris and other cell-related materials occupying the free surface and competitively reducing bead adsorption. As the bead-to-T cell ratio is critical for T cell activation or purification, the bead preparation step involving dilution with cell-free medium should be carefully considered to reach the desired bead concentration and ensure a consistent manufacturing process.

[0124] When using FIM, a feasible boundary approach is required for accurate quantification of specific particle populations found within highly heterogeneous samples. Output parameters by the instrument's operating software can be useful for identifying particles based on morphology, but can be prone to high error rates. The uniformity of the Dynabeads images resulted in similar values ​​of each particle parameter, and for Dynabeads-only samples, the developed filter had very little misclassification and 5% unclassified particles. However, the morphological nature of cells, cell aggregates and debris was very heterogeneous, and the distribution of each of these particle parameters was very dispersed (Figure 1). Furthermore, adducts and cells often had interchangeable values ​​for most particle parameters. The error rate for images containing only cells (misclassified and unclassified particles) was approximately 20% for the data set used to develop the morphological filter. The error rate for cells and Dynabeads suspensions was approximately 10% for the test data set. The lack of high power in discriminating between different particle population groups by using morphological parameters was one of the main motivations for using an automated image classification method based on CNN.

[0125] The high performance of neural networks is mostly based on large datasets to train the network. Millions of labeled images per class are required to successfully train an entire CNN such as VGG-19. Although FIM techniques are well suited for applications where a relatively large number of images can be collected in a relatively short period of time with low sample consumption, cleaning and labeling of high-quality training datasets remains a challenge. Since the pre-trained VGG-19 was able to efficiently distinguish features on the ImageNet dataset, fine-tuning the last two fully connected layers by using a relatively small number of labeled FlowCam images resulted in a powerful CNN for the discrimination and quantification of Dynabeads. The misclassification rate was significantly reduced with the CNN, being only 0.2% for cell samples spiked with Dynabeads. The FlowCam-CNN was able to quantify a wide concentration range of Dynabeads in cell suspensions, demonstrating a wide dynamic range. Moreover, the high accuracy (CV% less than 5%) in determining Dynabeads concentrations in cell suspensions above the LOQ presents the method as a robust approach for the quantification of process-related particles.

[0126] The determined total cell and debris concentrations were not affected by the number of Dynabeads added to the suspension. As expected, the number of detected adducts increased with increasing concentration of Dynabeads in the sample, due to the higher probability of the beads interacting with the cells. Furthermore, the concentration of 500,000 cells / mL showed no effect on the quantification of Dynabeads in the cell suspension. However, at the highest cell concentration tested, a clear decrease in the recovery of Dynabeads spiked to the target concentrations of 50,000 and 100,000 beads / mL was observed. This decrease in recovery was not observed at the lowest Dynabeads concentration of 20,000 beads / mL. A possible explanation could be the approach to count Dynabeads attached to a single cell. If a particle is classified as an "adduct", it is considered to consist of a single cell and a single Dynabead, which is the case in the majority of cases. However, as the number of cells or Dynabeads increases, the probability of capturing a cell with more than one attached Dynabead per image increases. Therefore, the underestimation of Dynabeads may have been related to inaccurate counting of beads in high-density cell populations.

[0127] The aforementioned examples teach a reliable method based on FIM combined with CNN for the detection, characterization and quantification of relevant particulate impurities. They show that small amounts of Dynabeads can be detected in cell suspensions when the bead concentration is above the determined LOQ, and high counting accuracy is achieved. Furthermore, cells and cell impurities, such as cell aggregates and adducts, can also be classified by using CNN. Quantification of these particles is useful for monitoring the manufacturing process of CBMPs and for process and product characterization, such as stability testing.

[0128] Materials and Methods for Examples 13-14-18 Unless otherwise stated, the materials and methods described below relate to Examples 14-18.

[0129] Cell and sample preparation. T-cell leukemia cells (Jurkat, clone E6-1, ATCC® TIB152™) were provided by Leiden University Medical Centre (LUMC) at 1×10 7 The cells were provided as frozen 1 mL aliquots at a total cell concentration of 1000 cells / mL (cells were counted using a NucleoCounter3000 [Chemometec]) and stored at -145°C prior to use. Jurkat cells were formulated in high glucose RPMI 1640 (RPMI medium; ThermoFisher, Waltham, USA) supplemented with 10% (v / v) fetal bovine serum (FBS; Life Technologies, USA) and 10% (v / v) dimethyl sulfoxide (DMSO) (Life Technologies, Carlsbad, USA). For washing and dilution steps, RPMI medium, 1x phosphate buffered saline (Gibco™, pH 7.4; ThermoFisher, Waltham, USA) or fetal bovine serum staining buffer (FisherScientific, New Hampshire, USA) were used. FBS (heat-inactivated, sterile filtered) and DMSO (suitable for hybridomas, ≥99.7%) were purchased from Sigma-Aldrich (St. Louis, USA). Dead Cell Apoptosis Kit (Annexin V, fluorescein isothiocyanate [FITC], and propidium iodide [PI]), calcein AM dye, and CyQUANT™ LDH kit were purchased from ThermoFisher, Waltham, USA.

[0130] Forced degradation tests Jurkat cells were exposed to three types of forced degradation tests as described below. Each of the three stress conditions was performed in duplicate on different days. Stressed samples were characterized by using orthogonal methods to assess the quality of the cell suspension.

[0131] Thawing stress A single 1 ml aliquot of frozen cells was transferred from storage at -145°C on dry ice to a water bath (GFL, Burgwedel, Germany) and thawed at 37°C for 1 min 30 s (default thawing conditions). For investigation of other thawing temperatures, cells were also thawed in a water bath at 25°C for 2 min 40 s and at 5°C for 4 min 25 s. Thawing in a water bath melted only the ice adhering to the walls of the cryovial. Furthermore, a 1 ml aliquot of the partially frozen cell suspension was transferred to 40 mL of RPMI medium equilibrated to the thawing temperature, and the remains of the ice nuclei were dissolved within a few seconds. After centrifugation of the cell suspension at 350xg for 10 min at 20°C, the supernatant was discarded and the pellet was resuspended in 5 ml of RPMI medium. Total cell concentration and viability were assessed by using a CASY counter (Bremen, Germany) and were determined to be 1x10 6 Dilution to cells / ml was performed by adding the required volume of RPMI medium (containing DMSO or FBS where indicated). Jurkat cell suspensions were therefore thawed at three defined temperatures.

[0132] Freeze-thaw stress: DMSO chilled on wet ice was slowly added to the cell suspension (in wet ice) to freeze 1x10 6 Target concentrations of 1%, 2.5%, 5% and 10% (v / v) were reached at cell concentrations in cells / ml. Additionally, five 1 ml aliquots were prepared in 1.8 ml Nalgene cryovials and frozen at -18°C for 3 hours. This amount of time was sufficient for complete freezing of the aliquots. Prior to analysis, samples were thawed at 37°C for 2 minutes, pooled and washed with fresh RPMI medium. Control samples were stored at 20°C for 3 hours and otherwise treated identically to the frozen-thawed samples.

[0133] Thawing Stress 1x10 6Cells at a concentration of 1000 cells / ml (with or without 10% [v / v] FBS) were filled (1 ml) into five Nalgene cryovials (ThermoScientific, Waltham, USA,) with a volume capacity of 1.8 ml. The cryovials were shaken by constant horizontal agitation at 185 rpm at 25°C by using an IC4000 shaker (IKA, Germany). Analysis of the cell suspension of the pooled cell aliquots was performed after 3 hours of shaking. Control samples were stored statically and protected from light at 25°C in cryovials and subsequently pooled.

[0134] Flow Imaging Microscopy (FlowCam, FIM) Cells (viable and necrotic) and debris particles were quantified by using a FlowCam 8100 (Fluid Imaging Technologies, Scarborough, USA) equipped with a 50 μm flow cell. The objective lens used provided a magnification of 20x and sample imaging was performed at 27 frames per second by a high-resolution CMOS camera (1920x1200 pixels). A total sample volume of 140 μL was measured with an efficiency of about 63% (i.e., the imaged sample volume was about 87 μL). Particles imaged within the flow cell were detected at an intensity threshold of 12 for bright and dark pixels. Washing steps between sample measurements included extensive washing of the flow cell with Terg-a-zyme® enzyme detergent (1% [w / v]) and highly purified water. Unless otherwise stated, collected images were not filtered using morphological filters and samples were measured in triplicate.

[0135] Image Analysis Using Convolutional Neural Networks Particle images captured using FlowCam were analyzed using a convolutional neural network (CNN). The CNN was based on the VGG-19 architecture described in Example 1. Briefly, a VGG-19 network using a modified linear unit activation function separated by max pooling and dropout layers was pre-trained on the ImageNet dataset (http: / / image-net.org). Retraining of the network was performed by freezing the first 10 layers to save computation time and improve the accuracy of the model by maintaining its pre-trained feature recognition ability. For fine-tuning, the dataset was split in a ratio of 8:1:1 for training, validation and testing, respectively. The model was optimized by running 19 epochs with the SGD optimization algorithm and a learning rate of 0.001. The machine learning model was implemented using Keras (2.2.4), Tensorflow (1.13.1) Python (3.7.3) libraries and was run on an Nvidia Turing GPU with 11 GB of VRAM.

[0136] Generation of particle images for population discrimination The CNN was retrained with manually labeled images (4,000-4,500) of each of the three populations: live cells, necrotic cells and debris particles. Images of live cells were manually selected from a measured sample consisting of Jurkat cells with a viability of 86.8% (based on the PI assay). For the selection of necrotic cells, two separate methods were used to induce cell necrosis. First, FlowCam images of a cell suspension treated with heat (55 °C for 90 min) were collected, which showed a viability of 7.6% (based on the PI assay). Second, cells were imaged after incubation with ethanol (10% (v / v)) for 90 min at 37 °C, with a determined viability of 12.8% (based on the PI assay). Morphological parameters such as aspect ratio, circle fit, convexity, sigma intensity and symmetry of the two populations were used to further define the image populations of live and necrotic cells for network fine-tuning. Only particles (cells) with morphological parameter values ​​falling between the 10th and 90th percentiles of the manually selected populations of live and necrotic cells were used as input for training the model. Images of debris particles were manually selected from measured cell suspensions subjected to vortexing for approximately 2 min or two freeze-thaw cycles (-140°C to 37°C).

[0137] Automated cell counting: A CASY TTC150 (Omni Life Science, Bremen, Germany) cell counter with a 150 μm capillary was used for automated cell counting. Cells were diluted 100-fold with CASYton (Omni Life Science, Bremen, Germany) immediately before analysis. Samples were measured in duplicate, each measurement consisted of 5 subruns of 400 μL, and more than 5,000 events were counted per analysis.

[0138] Imaging flow cytometry To assess cell viability based on cell metabolism, imaging flow cytometry was used in combination with the membrane-permeable dye calcein-AM (Calcein-AM assay) at a final concentration of 6.3 nM in the samples analyzed. After dye addition, cells were incubated for 30 min (in the dark) at room temperature and further washed with 200 μl of staining buffer (with FBS; BD Bioscience, New Jersey, USA). Samples were subsequently analyzed using an Amnis FlowSight imaging flow cytometer (Luminex, Seattle, USA) and data were analyzed using IDEAS6.2.183 (Luminex, Seattle, USA) image analysis software. Sample measurements were performed using a 20x magnification objective in medium sensitivity mode (132 mm / sec). The side scatter laser was set to 10 mW and the 488 nm fluorescence excitation laser was set to 12 mW. Images of Calcein-AM stained cells were recorded in channel 02 (532 / 55 nm). Brightfield acquisition was set on channel 9 (582 / 25 nm) with intensity set automatically. Only particles with a gradient RMS >30 were collected as true events. To assess cell viability only in the cell population, appropriate morphological filters were applied to distinguish between cells and debris particles. Cell viability was calculated based on the percentage of live cells from the total population (Figures 14A-14E). Each measurement was performed in duplicate.

[0139] Cell apoptosis and necrosis were determined by imaging flow cytometry after labeling cells with FITC and PI. Sample preparation was performed according to the Dead Cell Apoptosis Kit instructions. After 15 min incubation in the dark at room temperature, samples were kept on wet ice until analysis. The same instrument settings were used as described for the Calcein-AM assay above after adjusting the excitation power of the FlowSight imaging flow cytometer to 4 mW. Bright field as well as fluorescence images in channel 02 (532 / 55 nm) for FITC and channel 4 (610 / 30 nm) for PI were recorded. Appropriate morphological filters based on mean intensity, area, aspect ratio and adjustment from channel 09 were applied to distinguish single cells from cell aggregates or debris particles. Once single cells were selected based on the bright field images, the mean intensity values ​​from channel 02 (fluorescence) were plotted against the particle area (erosion mask) where positively stained cells were classified as live cells. Cell viability was calculated based on the percentage of live cells from the total population shown in Figures 14A-E.

[0140] Lactate dehydrogenase release assay. Plasma membrane integrity was assessed by using the lactate dehydrogenase (LDH) release assay (CyQUANT; ThermoFisher, Waltham, USA). Loss of cell integrity was measured based on the level of LDH released upon damage to the plasma membrane. This assay is based on the NADH + The assay is based on the conversion of lactate to pyruvate in the presence of LDH with the parallel formation of . Absorbance measurements were performed at 490 nm using a Tecan Spark plate reader (Tecan, Mannedorf, Switzerland) with a reference absorbance of 680 nm. The total cell concentration was adjusted to approximately 150,000 cells / mL with RPMI medium. (The cell concentration is within the linear range of this assay (data not shown). After incubation of the cells at 37°C and 5% CO2 for 45 min (with spiked water or Triton X-100), the assay was performed according to the manufacturer's protocol. The amount of LDH released in the medium was calculated according to Equation 3.

[0141] JPEG2024517592000003.jpg15154

[0142] Spontaneous LDH activity was based on measurements of cell suspensions spiked with water, and maximum LDH activity values ​​were derived from cell suspensions spiked with Triton X-100 to a target concentration of 10% (v / v). Results for each sample were based on measurements of four replicate wells.

[0143] Example 14 - FIM-CNN for measuring cell number, cell viability and debris particle concentration A CNN was trained for image classification using FlowCam images of live and necrotic cells as well as debris particles (Figure 15A). The recall value (proportion of positive identifications of all true positives present in the population) for classifying debris particles using the model was 99.5%, proving the high reliability of this classification approach (Figure 15B). Similar recall values ​​were achieved for live and necrotic cell populations (94.0% and 96.6%, respectively). To evaluate the performance of the classification model, the F1 score is an additional metric used to evaluate the weighted average of precision and recall. Here, the F1 scores for debris, live and necrotic cells were determined to be 0.991, 0.933 and 0.919, respectively. Such high values ​​confirm the low number of incorrectly classified images during the testing phase of retraining the network. Thus, in this study, FIM-CNN is presented as a novel method for determining the concentration of live and necrotic cells as well as for quantifying debris particles with heterogeneous size distribution (Figure 15C).

[0144] Example 15 - Effect of Thawing Temperature Frozen Jurkat cell suspensions were thawed at 5°C, 25°C and 37°C to assess the effect of thawing temperature on cell quality (Figure 16A). FIM-CNN and automated cell counting showed increased recovery of viable cells with increasing thawing temperature. Comparable concentrations of viable cells were obtained from both methods, but counts of necrotic cells and debris particles were substantially different. FIM-CNN measured more necrotic cells in samples thawed at 5°C, as well as an increased concentration of debris particles with decreasing thawing temperature. At the lowest tested thawing temperature of 5°C, cell suspensions contained 4x10 necrotic cells, including dark and dense particles that were not observed in cell suspensions thawed at 37°C and 20°C. 6 The debris particles contained more than 1 / mL (Figures 17A to 17B).

[0145] High cell viability was recorded for aliquots thawed at 37 °C and 20 °C (Figure 17B). The two fluorescence-based methods showed viabilities between 86% and 91%, compared to 63%-71% for FIM-CNN values. However, a clear decrease in cell viability to 43%, 50% and 22% was observed for aliquots thawed at 5 °C according to calcein-AM, PI and FIM-CNN assays, respectively. For all three thawing temperatures, there was a discrepancy in the absolute values ​​of determined cell viability when comparing the two fluorescence-based assays with FIM-CNN.

[0146] The Annexin V-FITC assay suggests that cells are more likely to undergo apoptosis when thawed at lower temperatures (Figure 16C). Notably, a significant increase in apoptotic cells can be observed in cell suspensions thawed at 5°C. A similar trend, although less pronounced, was observed in the LDH release assay (Figure 16D).

[0147] Example 16 - Effect of Freezing and Thawing Jurkat cells formulated in RPMI medium and DMSO (0%, 1%, 2.5%, 5% and 10% [v / v]) were subjected to one freeze-thaw cycle involving storing the cell suspension at −18° C. for 3 h and thawing at 37° C. for 2 min. Freezing and thawing the cell suspensions containing 0% and 1% (v / v) DMSO resulted in a greater than 98% reduction in cell number, indicating that virtually all cells lose their viability under these conditions (data not shown).

[0148] The concentrations of cells (viable and necrotic) and debris particles were determined using FIM-CNN (Figures 18A-18B). At higher DMSO concentrations, higher total cell concentrations were observed after one freeze-thaw cycle compared to the lower tested cryoprotectant concentrations. Furthermore, cell suspensions containing 10% (v / v) DMSO contained the lowest number of debris particles after freeze-thaw. Total cell concentrations before freezing (T0) and after 3 hours of storage at room temperature (control) were comparable for each tested DMSO concentration. However, the control cell suspensions containing 10% (v / v) DMSO showed a slightly higher debris content compared to T0 and cell suspensions with lower DMSO content.

[0149] Interestingly, cell viability determined after freeze-thaw was highly dependent on the assay used (Figure 18C). The calcein-AM assay suggested minimal changes in cell viability, with 85-90% cell viability, regardless of DMSO content and time point. In contrast, the PI assay and FIM-CNN assay showed a notable (approximately 20%) loss of viability for cell suspensions frozen with 2.5% (v / v) DMSO, indicating that this concentration of cryoprotectant is insufficient for cell stabilization upon freezing. The PI assay also showed smaller drops of cell viability of approximately 10% for cell suspensions containing 5% and 10% (v / v) DMSO for freeze-thawed and control samples compared to T0, whereas the FIM showed such a drop in cell viability only for cell preparations containing 10% (v / v) DMSO. The toxic effect of DMSO was demonstrated by the loss of cell viability at 10% (v / v) based on the FIM after a 3 h exposure at room temperature for each tested concentration according to the PI assay. Based on the results from PI and FIM, the best cell stabilizing properties of DMSO were achieved at 5% (v / v).

[0150] An increase in LDH release in the medium was observed in cell suspensions that had undergone freeze-thawing or were stored at room temperature in the presence of DMSO (Figure 19A). Only slight differences in LDH concentrations were observed in cell preparations with different DMSO concentrations. However, cells frozen in 5% (v / v) DMSO showed less LDH release compared to 2.5% and 10% (v / v) suspensions. Similarly, the percentage of apoptotic cells was minimal in 5% (v / v) cell suspensions upon freeze-thawing (Figure 19B).

[0151] Example 17 - Effect of Shaking Stress Jurkat cell suspensions were shaken horizontally in the presence and absence of 10% (v / v) FBS. Figure 20A shows the total concentration of live and necrotic cells and debris particles measured by FIM-CNN in the two cell preparations at a given time point. After shaking stress, a substantial decrease in the total number of cells was observed in the preparation without FBS compared to the cell suspension containing FBS. For both preparations, a significant increase in the concentration of debris particles was observed. Notably, the debris present in the stressed cell suspension supplemented with FBS was derived not only from ruptured cells but also from aggregated proteins present in the FBS (observed in stressed RPMI medium containing 10% [v / v] FBS, data not shown). Cell viability after shaking stress and resting storage was very similar to the viability of cells at T0 by all three methods used (Figure 21).

[0152] LDH release in stressed cell suspensions was higher compared to T0 or control cell suspensions, regardless of FBS content. The increase in LDH release in formulations containing FBS was unexpected, given the observation that cell viability and cell concentration remained unchanged at all sampling points. A slight increase in the apoptotic cell fraction was observed after shaking formulations with and without FBS, while a similar increase was observed in the non-shaken control. Thus, shaking stress had no detectable effect on cell apoptosis in this study.

[0153] Example 18 - Summary of Examples 13-17 Currently approved CBMPs are stored either in cryopreserved (frozen liquid) or non-frozen (liquid) state. In both cases, cells can be exposed to all kinds of stress factors, for example resulting from freezing-thawing or handling in the clinic before administration. This can cause cell damage, loss of quality and potential clinical relevance, for example severe adverse effects or lack of drug efficacy. Cells have developed mechanisms to handle a certain number of stress stimuli by activating signaling pathways and stress response proteins. However, if a threshold of harmful factors is exceeded, cells can suddenly lose membrane integrity or undergo programmed death. In the human body, macrophages and dendritic cells of the innate immune system are equipped with appropriate mechanisms to remove dead cells. However, in CBMPs, any cellular decomposition that may be formed, such as dead cells and cell debris, may remain in the product and accelerate further cell degradation. Moreover, debris particles derived from suddenly destroyed cells lack the "eat-me" signal of apoptotic cells. After administration, such necrotic cells and debris particles can trigger the immune system potentially resulting in an inflammatory response.

[0154] Viable cell concentration is considered one of the most important quality attributes of CBMPs. A variety of viability assays based on cell membrane or cell metabolic integrity are available. However, common viability assays involve fluorescent staining that involve expensive fluorescent dyes, multi-step preparation procedures with incubation times, or interference with medium components. Furthermore, manual gating in flow cytometers is highly biased and difficult to reproduce between different operators. Thus, the shortcomings of these assays limit their ability to be robust and high-throughput methods for quantification of necrotic cells. Moreover, classical trypan blue exclusion and other colorimetric assays are characterized by low precision and often show cytotoxic effects on cells.

[0155] In the study disclosed herein, FIM-CNN was applied to accurately determine the number of live and necrotic cells in suspensions of up to 200,000 cells / mL based on the morphological appearance of the cells. Necrosis is manifested by several morphological changes to the appearance of the cells, including cell rounding and swelling of the cytoplasm. These are not evident in FlowCam images relative to the human eye and are difficult to distinguish using the morphological parameters output by the FlowCam software. However, the pattern recognition algorithm implemented in the CNN model has been shown to successfully distinguish complex fingerprints, i.e., protein aggregate structures formed in different stress conditions, and proved successful in classifying cell-associated particles in this study. It should be noted that each of the viability assays used may yield different results, considering the different basis on which viability is assessed. For example, when using the calcein-AM assay, some apoptotic cells are metabolically active and show positive staining, while others may be classified as necrotic. With the Annexin V-FITC and PI assays, only PI-positive cells were classified as necrotic cells, while cells that underwent early apoptosis were considered viable. Under the conditions used in these experiments, it was not possible to discriminate apoptotic cells using FIM-CNN, so apoptotic cells were randomly classified as either live or necrotic cells.

[0156] Characterization of particulate matter in parenterals such as CBMPs is described in pharmacopoeial monographs, and products should meet acceptable limits for subvisible particulate impurities with sizes above 10 μm and 25 μm. However, given the particulate nature of cells, it is difficult to meet specific particle testing requirements applicable to injectable products, as human cells fall within the subvisible size range (typically 7-30 μm). Light obscuration, one of the standard pharmacopoeial methods for quantification of subvisible particulate impurities, does not distinguish between cells and other particulates. High-throughput microscopy methods that provide morphological data on particles within hundreds of μl are therefore promising tools for evaluating particulates in cell-based products.

[0157] Here, by using FIM-CNN, we were able to obtain reliable counts for the concentration of debris particles with sizes between 1 and 50 µm. Debris particles can be considered as particulate impurities in CBMPs, as they consist of a mixture of lipids, proteins, nucleic acids, and potentially other exogenous particles. Nevertheless, commonly applied cell characterization methods do not report the debris content within cell suspensions, likely omitting relevant information regarding their potential impact on the product quality (consistency and stability) and safety of these drug products. Moreover, debris content imaged by using bright-field microscopy in cell suspensions was previously used to assess the levels of late apoptotic primary human hepatocytes and MDCK cells. Similarly, it was observed that exposure of cells to stress conditions led to an increase in cellular debris that correlated well with a decrease in cell viability and total cell number.

[0158] The proprietary supply chain of some CBMPs, which includes at least one freeze-thaw cycle, makes cryopreservation very important in preserving the integrity of the product. Cryopreservation of human cells has been studied extensively, and multiple factors have been found to be critical in the process. Freezing and thawing rates are two of the most important parameters that should be optimized to achieve high recovery of functional cells after thawing. However, optimal conditions vary for each cell suspension, and the effect of thawing rate is less understood compared to the effect of freezing rate on cells. Jurkat cells used in this study were frozen at 1 °C / min with 10% (v / v) DMSO. -1The cells were then subjected to controlled freezing at a rate of 100 µg / ml. Subsequent thawing at 5 °C in a water bath had a negative effect on cell viability and apoptosis rate. The adverse effect of slow thawing on cells may be due to ice recrystallization. Transitional ice crystals formed during freezing may form larger crystals upon thawing at 5 °C, leading to denaturation of cellular proteins and disruption of membranes. A similar observation was made by Thorpe et al., who reported that decreasing the thawing temperature reduced the viability of mouse lymphocytes. On the other hand, the results presented herein are not in agreement with previous studies, where thawing temperature had a minimal effect on T cell viability. In addition to the reduced viability, an increased concentration of debris particles was observed at lower thawing temperatures, with dense and irregular particles observed at the lowest thawing temperature. Such debris particles were not observed in cell suspensions thawed at 20 °C and 37 °C.

[0159] The presence of a cryoprotectant is crucial for the cryopreservation of cells stored at ultra-low temperatures (below 120 °C in the gas phase of liquid nitrogen). DMSO is currently the most widely used cryoprotectant in cryobiology. Its amphipathic and water-binding properties allow it to easily cross cell membranes, thereby avoiding the outflow of water from the cytoplasm and thus preventing cell dehydration upon freezing. At commonly used DMSO concentrations such as 5-10% (v / v), the formation of large intra- and extracellular crystal lattices is prevented by interfering with water molecules. The formation of ice crystals and the maintenance of cell viability also depend on the cooling rate, with very slow cooling rates (0.1 °C / min) being preferred. -1 ) produces fine dendritic ice structures and a fast cooling rate (10°C / min -1) produces large ice crystals. In this study, Jurkat cells were subjected to passive (uncontrolled) freezing at -18 °C in the presence of DMSO at concentrations ranging from 0% to 10% to mimic unintentional freezing. The lowest tested DMSO concentrations (i.e., 0% and 1%) did not show any cryoprotection to the cells upon freeze-thaw, as an almost complete loss of cell viability was observed. In contrast, the highest tested DMSO concentration of 10% (v / v) showed the highest cell recovery after one freeze-thaw cycle and the least amount of debris particles were present in the cell suspension. However, the control sample (stored at room temperature) with the highest tested DMSO concentration showed an increase in debris particles over time, most likely due to its cytotoxic effect on the cells.

[0160] Unexpectedly, the three viability assays used in this study did not provide conclusive results on the degree of protection against freeze-thaw at each tested DMSO concentration. The calcein-AM assay showed the least sensitivity in detecting changes in cell viability after freeze-thawing, whereas the PI and FIM assays detected a substantial decrease in viability in suspensions containing 2.5% (v / v) DMSO. However, for cell suspensions containing 5% and 10% (v / v) DMSO, a decrease in viability was observed in the PI assay for freeze-thawed and control samples, whereas the FIM suggested that freeze-thaw had a minimal effect on cell viability at 5% (v / v) DMSO. Different assays may differ in their level of sensitivity to cell viability due to their different measurement principles. Moreover, post-thawed cell suspensions containing 5% (v / v) DMSO showed the smallest fraction of apoptotic cells (based on Annexin V-FITC assay) and the lowest LDH release compared to the other two tested DMSO concentrations. Similar observations were made for peripheral blood stem cells. After cryopreservation with 5% DMSO, fewer apoptotic and necrotic cells were measured compared to suspensions containing 10% DMSO. The apoptosis-inducing effect of DMSO through its interaction with PD-1 receptors and their ligands was observed in the control group, where an almost two-fold increase in the apoptotic cell content was recorded.

[0161] Mechanical stress may be the most frequent stressor to which biopharmaceuticals are exposed during processing and handling. Solid-liquid and air-liquid interfaces play an important role in the formation of particles in protein-based formulations during agitation. Thereby, amphiphilic non-ionic surfactants (e.g., polysorbate 20 or 80) are commonly used to reduce interfacial stress and protein aggregation upon mechanical shock. Unfortunately, such an approach may not be successful for CBMPs, since polysorbates have a lytic effect on cells even at low concentrations. To date, mild shaking conditions have been applied to cell suspensions mainly for cell proliferation purposes or to induce the formation of cell aggregates to create ex vivo tissue models. In this study, more severe stress conditions were introduced to mimic potential "real-life" mechanical stresses that have been shown to have detrimental effects on protein-based therapeutics. Cell suspensions filled in cryovials and positioned vertically were not affected by shaking speeds up to 500 rpm. However, placing the cryovials horizontally led to a dramatic (approximately 50%) loss of total cell concentration after a relatively short shaking stress at 185 rpm. Cell viability was not affected by mechanical stress by the two fluorescence-based methods and FIM-CNN, whereas LDH release and apoptosis were significantly increased compared to control samples. The increase in free LDH in the medium suggests a decrease in cell quality and indicates the value of using orthogonal cell characterization techniques. Furthermore, FBS showed a protective effect on cells upon shaking. FBS is a commonly used supplement in cell culture that may protect cells from mechanical stress, but is considered a process-related impurity that is difficult to remove in downstream processing steps. Furthermore, FBS contains proteins that are susceptible to aggregation and particle formation upon mechanical stress. Protein aggregates may induce unwanted immunogenicity in themselves, which adds another undesired and overlooked complexity with respect to particulate impurities in CBMPs. Ultimately, a broader range of compounds should be tested in cell suspensions subjected to forced degradation studies to enable decision-based formulation strategies for CBMPs.

[0162] Herein, the application of systematic forced degradation testing in the evaluation of cell stability and formulations is disclosed, using Jurkat cells as a model and FIM-CNN to determine cell damage. A homogeneous cell line of immortalized cells may not fully reflect the behavior of cell samples used in the clinic. Aging of primary human cells introduces much greater heterogeneity in morphological cell features, resulting in a greater challenge for the acquisition of quality parameters. However, the application of CNN for image classification allows the utilization of FIM as a robust and rapid analytical tool for the characterization of cell suspensions, which may have potential applications in the evaluation of "real-life" samples. In addition to other established approaches, cell viability was assessed by using FIM-CNN, and good correlation between the methods was observed. Furthermore, quantification of debris particles in CBMPs was only possible with FIM-CNN, highlighting the ability of the method to monitor the levels of particulate impurities in cell suspensions. Overall, this study shows that thermal stress, freeze-thaw stress and osmotic stress are relevant tools for formulation studies on CBMPs. Furthermore, the importance of selecting a wide range of analytical methods to better understand the impact of each stressor on the quality of CBMPs is presented herein.

[0163] Example 19 - Identification of apoptotic cells The following materials and methods were used to identify live, apoptotic and necrotic cells and debris particles.

[0164] Flow Imaging Microscopy (FlowCam, FIM) Cells (viable, apoptotic, and necrotic) and debris particles were quantified by using a FlowCam 8100 (Fluid Imaging Technologies, Scarborough, USA) equipped with a 50 μm flow cell. The objective lens used provided a 20x magnification, and sample imaging was performed at 27 frames per second by a high-resolution CMOS camera (1920x1200 pixels). A total sample volume of 140 μL was measured with an efficiency of about 63% (i.e., the imaged sample volume was about 87 μL). Particles imaged within the flow cell were detected with a threshold of 12 for bright and dark pixels. Washing steps between sample measurements included extensive washing of the flow cell with Terg-a-zyme® enzyme detergent (1% [w / v]) and highly purified water. Unless otherwise stated, collected images were not filtered using morphological filters, and samples were measured in triplicate.

[0165] Image Analysis Using Convolutional Neural Networks

[0166] Particle images captured using FlowCam were analyzed using a convolutional neural network (CNN). The CNN was based on the VGG-19 architecture described in Example 1. Briefly, a VGG-19 network using a modified linear unit activation function separated by max pooling and dropout layers was pre-trained on the ImageNet dataset. Retraining of the network was performed by freezing the first 10 layers to save computation time and improve the accuracy of the model by maintaining its pre-trained feature recognition ability. For fine-tuning, the dataset was split in a ratio of 8:1:1 for training, validation and testing, respectively. The model was optimized by running 15 epochs with the SGD optimization algorithm and a learning rate of 0.0005. The machine learning model was implemented using Keras (2.2.4), Tensorflow (1.13.1) Python (3.7.3) libraries and was run on an Nvidia Turing GPU with 11 GB of VRAM.

[0167] Generating particle images for population identification

[0168] The CNN was retrained with manually labeled images (4,000–4,500) of each of the four populations: live cells, apoptotic cells, necrotic cells, and debris particles.

[0169] Images of live cells were manually selected from a measurement sample consisting of freshly collected Jurkat cells with a viability of 91.6% (based on the PI assay). Image selection was performed by applying a size filter (10-20 μm) and an aspect ratio filter (0.6-1.0). Furthermore, image selection was manually refined by excluding visible apoptotic blebs and cells with rougher surfaces. To generate images of apoptotic cells, Jurkat cells were treated for 3 h with staurosporine (1 μM-5x10 5Cells). The apoptotic rate was approximately 90-95%, as confirmed by flow cytometry (Annexin V-FITC). A size filter (10-20 μm) was applied to exclude debris and aggregates, and a circularity filter (less than 0.9) was used to exclude the remaining live cell images. Cell selection was manually double-checked to exclude some debris particles. For the selection of necrotic cells, two separate methods were used to induce cell necrosis. First, FlowCam images of cell suspensions treated with heat (55 °C for 90 min) were collected, which showed a viability of 7.6% (based on PI assay). Next, cells were imaged after incubation with ethanol (10% (v / v)) for 90 min at 37 °C, with a determined viability of 12.8% (based on PI assay). Morphological parameters such as population aspect ratio, circle fit, convexity, sigma intensity and symmetry were used to further define the image population of necrotic cells for network retraining. Only particles (cells) with morphological parameter values ​​falling between the 10th and 90th percentiles of the manually selected population of necrotic cells were used as input for training the model. Images of debris particles were manually selected from measured suspensions of cells subjected to vortexing for approximately 2 min or two freeze-thaw cycles (-140°C to 37°C).

[0170] Example 20 - Generation of images of apoptotic cells (1) Methods: Cells were treated with camptothecin (8 μM, 14 h at 37° C. and 5% CO2). Images of treated cells were manually selected. The CNN was retrained with manually selected live, necrotic, apoptotic cells and debris particles.

[0171] Results When the apoptotic cell population was included, the classification accuracy was low - 78.4% (Figure 22). The accuracy of the model was insufficient to detect changes in apoptosis over time (Figure 23).

[0172] Example 21 - Generation of images of apoptotic cells (2) The goal of this experiment was to generate a library of cell images for each class with few or no cell images belonging to other classes, e.g., minimizing necrotic cells in an apoptotic sample.

[0173] Methods To generate apoptotic cells, the reagent staurosporine was used (1 μM-500.000 cells / mL) and the apoptosis rate was confirmed by a fluorescence-based assay using a flow cytometer. Double staining with Annexin V-FITC & PI was used to distinguish between live, apoptotic and necrotic cells. The apoptosis rate was about 90%, live cells about 2% and necrotic cells about 6%. Furthermore, a large amount of debris particles was present in the samples, which was excluded in the flow cytometry analysis by gating by forward and side scatter. The apoptotic samples were measured in parallel with the FlowCam instrument (triplicate measurements). By applying size filters (area-based diameter 6-25 μm), circularity filters (0.5-1) and aspect ratios (0.6-1), the majority of debris particles could be easily excluded. The exclusion of the remaining debris particles as well as healthy and necrotic cells is a bigger challenge due to the amount of cell images (90,000 per replicate) and the close similarity of their morphological appearance by the naked eye.

[0174] Freshly collected cells were measured by FlowCam. Reference measurements by using a fluorescence-based flow cytometry assay confirmed high cell health with over 92% live cells, approximately 3% apoptotic cells, and 5% necrotic cells. Cell debris was excluded by using the filters described above, and cell images were also manually double-checked by the operator. All cell images similar to images from apoptotic or necrotic samples were excluded. Additionally, all cell images that were not completely in focus and were round were also excluded.

[0175] Results An example image is shown in Figure 24. A new CNN model was trained using this new dataset and the training results are shown in Figure 25.

[0176] When samples were tested with the resulting CNN, an underestimation of cell viability occurred, showing a deviation relative to the fluorescence-based flow cytometry assay of approximately 20-25% (Figure 26). Manual reselection of healthy cells may have introduced bias in the CNN, with out-of-focus cells being classified as debris, necrotic or apoptotic, but not as healthy cells.

[0177] To check to what extent overselection biased the CNN, a pre-filter on edge gradient (a parameter of focus) was tested. In detail, images obtained from FlowCam measurements were filtered before CNN analysis with an edge gradient filter (50-70) and only the remaining focused cell images were tested with the CNN model. The results obtained by CNN were in better agreement with the An.V-FITC assay (a fluorescence-based flow cytometry assay) with a lower deviation of about 10-15%.

[0178] When the out-of-focus cell images were removed from the test dataset, the CNN showed relatively more live cells (and fewer apoptotic cells) (Figure 27), suggesting that the out-of-focus live cells were misclassified as apoptotic cells.

[0179] Example 22 - Use of fluorescent signals to select training images The goal of this experiment was to identify necrotic, apoptotic and live cells by their fluorescent signals and train a CNN using brightfield images of the cells.

[0180] Methods The CNN described in the previous example was trained on brightfield images of necrotic, apoptotic and live cells. Brightfield images were selected based on coincident fluorescent signals in the instrument's fluorescent channels. Using this approach, necrotic cells (positive PI signal), apoptotic cells (positive Annexin V-FITC signal) and live cells (positive Calcein-AM signal) can be individually identified and subsequently used to train the CNN.

[0181] Two cell samples were stained with calcein-AM to determine viable or necrotic cells, and two samples were stained with Annexin V-FITC and PI to determine apoptosis and necrosis.

[0182] Samples were analyzed using an Amnis FlowSight imaging flow cytometer (Luminex, Seattle, USA) and data were analyzed using IDEAS 6.2.183 (Luminex, Seattle, USA) image analysis software. Sample measurements were performed using a 20x magnification objective in medium sensitivity mode (132 mm / sec). Images of calcein-AM stained cells were recorded in fluorescence channel 02 (532 / 55 nm). The 488 nm fluorescence excitation laser was set to 12 mW. Brightfield acquisition was set to channel 9 (582 / 25 nm) with an automatically set intensity. For Annexin V-FITC / PI, the excitation power of the FlowSight imaging flow cytometer was reduced to 4 mW and the same instrument settings as described above for the calcein-AM assay were used. Brightfield as well as fluorescence images were recorded in channel 02 (532 / 55 nm) for Annexin V-FITC and channel 04 (610 / 30 nm) for PI.

[0183] For training the CNN, for example, 4000 bright field images of the cells used were taken from the population. For necrotic cells, images were taken from the bright field channel corresponding to a positive signal in the fluorescent channel detecting PI. For apoptotic cells, images were taken from the bright field corresponding to a positive fluorescent staining with Annexin V-FITC. For live cells, images were taken from the bright field channel corresponding to a positive signal in the fluorescent channel detecting Calcein-AM. In this experiment, debris particles were excluded from the analysis.

[0184] Results Results from differentiation between live and necrotic cells by using calcein-AM staining are shown in Figure 28. For cell samples treated with digitonin to induce necrosis, there was good agreement between the results using brightfield images and CNN using fluorescence (Figure 28A). For freshly thawed cell samples, the CNN overestimated the number of necrotic cells compared to the fluorescence standard (Figure 28B).

[0185] When apoptosis was induced with camptothecin, CNN overestimated the population of apoptotic cells (Figure 29A). Furthermore, apparently necrotic and live cells were identified as apoptotic by CNN using brightfield images compared to a fluorescent standard (Figure 29A). Similarly, in the untreated negative control, CNN underestimated the number of live cells (Figure 29B).

[0186] FIG. 30 shows the confusion matrix for classification of live, necrotic and apoptotic cells.

[0187] Example 23 - Robustness of necrosis induction and CNN retraining The goal of this experiment was to generate a new dataset of necrotic cell images and retrain a new CNN to demonstrate the robustness of inducing necrosis and retraining the CNN.

[0188] Methods The CNN was retrained with a newly selected training image library according to the method described in Example 21. To do so, the generated cell images for live cells, apoptotic cells and cell debris were reviewed and 4,000-5,000 images were manually selected. For the necrotic class, a new dataset was generated by incubating cells at 55°C for 90 min or inducing necrosis by incubating cells with 10% (v / v) ethanol for 90 min. Approximately 4,000-5,000 necrotic cell images were manually selected.

[0189] Results Retraining the new CNN with a new dataset of necrotic cells obtained by the two different methods resulted in good classification accuracy for all four classes (Figure 31), comparable to the results from the previous CNN retraining (Figure 25). We then compared the distribution of predicted classes obtained by the CNN trained on the two different datasets of necrotic cells with the distribution of classes obtained by fluorescence-based flow cytometry by using images from measurements that were not part of the training. The CNN trained on the new set of necrotic cells was more accurate in predicting healthy cells (Figure 32), demonstrating the robustness of the CNN retraining and the optimized training dataset.

[0190] Example 24 - Generation of living cells with different morphologies The aim of this study was to train a CNN to discriminate between viable (healthy) cells with different morphologies. To do so, cells were suspended in solutions with different osmolalities, thus affecting the cell volume and therefore the cell morphology.

[0191] Methods Solutions were prepared in the hypotonic range by diluting phosphate buffered saline (PBS) with highly purified water. The osmolalities were 220, 250 and 270 mOsm / kg. Additionally, hypertonic PBS solutions were prepared by adding different amounts of sucrose. The osmolalities were 350, 400, 420 and 450 mOsm / kg. For isotonic conditions, PBS was used as provided by the manufacturer. Freshly harvested cells were used to create an image training library of live cells in each osmolality range, i.e., hypotonic, isotonic and hypertonic. Cell viability of isotonic samples was confirmed by Annexin V-FITC / PI and Calcein AM / PI fluorescence assays. The Annexin V-FITC / PI assay measured 77% live cells, 9% apoptotic cells and 14% necrotic cells. The Calcein AM / PI assay measured 89% live cells and 11% necrotic cells. Cells were centrifuged and resuspended in each solution at adjusted osmolarity for 5 min prior to FlowCam measurements. Live cell images were manually selected in the size range between 6 and 25 μm (diameter (ABD)) for each osmolarity range (4,000-5,000 images). The CNN was trained on these new live cell images, whereas the training library of apoptotic, necrotic cells and cell debris was the one used in Example 23.

[0192] Results Healthy cells were divided into "Hypo", "Iso" and "Hyper" for each osmolarity range. This avoided heterogeneity within the training images of live cells due to different cell morphologies in hypotonic (swelling) and hypertonic (shrinking) solution conditions. The CNN was trained on manually selected images (4,000-4,500) from each of six categories: cells in hypotonic solution, cells in isotonic solution, cells in hypertonic solution, apoptotic cells, necrotic cells and debris particles.

[0193] The CNN trained on the six cell categories showed high classification accuracy (Figure 33). The classes Iso, Hypo, and Hyper were then combined to determine the total amount of predicted live cells. The CNN had a distribution of predicted classes similar to the distribution of classes obtained with fluorescence-based flow cytometry (Figure 34) and similar to previous models 1-01 and 1-02 (Figure 32).

[0194] Subsequently, healthy cells were incubated for 5 min at osmolarities ranging from 220 mOsm / kg to 450 mOsm / kg, and their images were captured by FIM and then classified by CNN. The CNN successfully classified the cells according to the osmolarity conditions (Hyper, Iso, Hypo) in which they were incubated. As expected, the number of live cells in the "Hypo" class increased as the osmolarity decreased (hypotonic range below 300 mOsm / kg), while the number of cells in the "Hyper" class increased with increasing osmolarity (hypertonic range above 300 mOsm / kg) (Figures 35A and 35B).

Claims

1. 1. A method for quantifying single viable cells and / or particulate impurities in a cell-based product sample, comprising: a) providing a training set of classified flow-imaging microscopy (FIM) images including single live cells, debris, and at least one of necrotic cells, apoptotic cells, doublet cells, magnetic beads, or adducts; b) training a convolutional neural network (CNN) using the classified images of step (a); and c) capturing a FIM image from said product sample; d) applying the trained CNN obtained in step (b) to the FIM image of step (c) to identify single live cells, debris, and at least one of necrotic cells, apoptotic cells, doublet cells, magnetic beads, or adducts; e) quantitating the single viable cells and / or the particulate impurities identified in step (d); 1. A method for quantifying single viable cells and / or particulate impurities in a cell-based product sample, comprising:

2. The set of classified images is Single live cells, necrotic cells, apoptotic cells, doublet cells, magnetic beads, adducts and debris; Single live cells, doublet cells, magnetic beads, adducts and debris; Single live cells, necrotic cells and debris; or Single live cells, necrotic cells, apoptotic cells, and debris The method of claim 1 , comprising:

3. The classified images of step (a) i) providing a cell suspension; ii) applying said cell suspension conditions that induce necrosis, apoptosis, cell debris, adducts, or any combination thereof; iii) optionally labeling the cells; and iv) acquiring a FIM image of the cell suspension; v) classifying said FIM images; The method of claim 1 , wherein the compound is obtained by a method comprising:

4. said culture conditions for inducing necrosis and / or inducing debris comprising subjecting said cell culture to stress conditions; The culture conditions for inducing apoptosis include incubating the cell culture with a pro-apoptotic agent; the culture conditions for inducing adducts include incubating the cells with magnetic beads; or The method of claim 3, any combination thereof.

5. 5. The method of claim 4, wherein the stress conditions for inducing necrosis comprise mechanical stress, shaking stress, freeze-thaw, heat, incubation with ethanol, or any combination thereof.

6. The method of claim 4 , wherein the pro-apoptotic agent comprises staurosporine.

7. The method of claim 4, wherein a portion of the cells is incubated with the magnetic beads in a ratio of about 1:

1.

8. i) the conditions that induce necrosis include incubating the cells in cell culture medium containing 10% ethanol at 55° C. for 90 minutes or at 37° C. for 90 minutes; ii) the conditions for inducing apoptosis include incubating the cells with 1 μM staurosporine for 3 hours at 37° C.; iii) the debris-inducing conditions include vortexing the cells for 2 minutes in a 2 mL microfuge tube or subjecting the cells to two freeze-thaw cycles from −140° C. to 37° C.; iv) the conditions inducing adducts include incubating magnetic beads with cells in a 1:1 ratio at 37° C. for 1.5 hours; The method according to claim 3.

9. 2. The method of claim 1, wherein the necrotic cells from the training set include cells incubated in cell culture containing 10% ethanol at 37°C for 90 minutes and cells incubated at 55°C for 90 minutes.

10. 2. The method of claim 1, wherein the live cells from the training set include cells incubated in a hypotonic medium, cells incubated in an isotonic medium, and cells incubated in a hypertonic medium.

11. The classified images of step (a) i) obtaining a sample of said cell-based product; ii) optionally labelling said cells of said sample of step (i); iii) acquiring a FIM image of the sample of step (i) or (ii); iv) classifying the FIM images; The method of claim 1 , comprising:

12. The CNN pre-trained prior to step (b) using a pre-training set of images; Includes the VGG-19 architecture, or During the training of step (b), the last two fully connected layers are active or the first ten layers are frozen; or including combinations thereof, The method of claim 1.

13. The method of claim 1 , wherein the training in step (b) comprises an Adam optimization algorithm or an SGD optimization algorithm.

14. The method of claim 1 , wherein the cell product comprises T cells.

15. 1. A method for quantifying single viable cells and / or particulate impurities in a cell-based product sample, comprising: a) providing a cell suspension comprising single viable, necrotic and / or apoptotic cells; b) simultaneously obtaining bright field images and fluorescent signals from said cell suspension using an imaging flow cytometer, the fluorescent signal is used to assign a classification label to the brightfield image according to whether the brightfield image contains live cells, necrotic cells, or apoptotic cells; and c) training a convolutional neural network (CNN) using the bright-field images of step (b); and d) capturing a flow imaging microscopy (FIM) image from said product sample; and e) applying the trained CNN obtained in step (c) to the FIM image of step (d) to identify single live cells, debris, and at least one of necrotic cells, apoptotic cells, doublet cells, magnetic beads, or adducts; f) quantitating the single viable cells and / or the particulate impurities identified in step (e); 1. A method for quantifying single viable cells and / or particulate impurities in a cell-based product sample, comprising:

Images