Cell-based medicinal products (CBMPs) are a promising new class of biopharmaceuticals for the treatment of severe diseases such as cancer or genetic diseases [1]. Most commonly used cell types in clinical development are currently T cells, followed by stem cells and dendritic cells [2]. Also, six chimeric antigen receptor (CAR) T cells products are approved by the European Medicine Agency [3]. CAR T cells are genetically modified T cells to target cancer cells [4], whereas the CAR consists of an antigen-recognition domain fused to the T cell receptor [5]. The success of CBMPs depends on their effective manufacturing, which has an impact on product quality, patient safety and therapy effectiveness. CBMPs belong to the most complex biopharmaceutical products. They consist of living cells, which are either derived from the patient, i.e., autologous therapy, or are obtained from a donor, i.e., allogenic therapy [6]. The production process of CAR T cells is complex with several manufacturing steps from leukapheresis, transduction, expansion to fill-finish and distribution [1]. Fig. 1 shows a schematic overview of a CAR T cell production process from leukapheresis to the CAR T cell infusion. Within this process up to two cryopreservation processes may be applied, i.e., for the starting cell material after leukapheresis and for the final product [7]. Additionally, short-term liquid storage of the starting material is done up to 2–4 days at 2–8 °C [7], [8].

Throughout each processing step cells may potentially be damaged leading to a decrease in cell viability and an increase in cell related impurities. Obtaining a high viability is critical to ensure a potent and safe drug product. Outside cell culture conditions the cell survival is limited to several hours to days depending upon the cell type and storage conditions [9], [10]. Following CAR T cell production, the cells are usually cryopreserved with 5 to 10% dimethyl sulfoxide (DMSO) for frozen storage [7]. Although DMSO is a very effective cryoprotectant it has detrimental effects on cell viability at ambient temperatures [11]. Therefore, the exposure time of cells to DMSO must be kept to a minimum [11]. However, during the production process hold-times in cryoprotectant solution may occur, e.g., due to sample batch size and a prolonged filling procedure [12]. The effect of DMSO on cells can depend upon the used cell type, the storage condition (e.g., time and temperature), and the concentration of DMSO [11], [13]. Additionally, the choice of the cell culture medium can influence the cell viability [14]. For T cells RPMI medium is commonly used for cell culture, however PBS is a standard solution for some handling steps such as washing for cell staining. Furthermore, some buffer systems are known to lead to changes in pH upon freezing such as phosphate buffers [15], which may influence the cell viability throughout cryopreservation. The implications of short-term liquid storage during CBMP processing on cell viability need to be explored further to define accurate limits for holding times.

To evaluate cell viability, fluorescence-based assays are most commonly used in combination with flow cytometry [16]. Additionally, machine learning tools are increasingly explored for the analysis of cells [17], [18]. Previously, flow imaging microscopy (FIM) in combination with a convolutional neural network (CNN) was used to differentiate between viable and dead cells without any additional sample preparation such as staining [19]. Similarly, Thite and coworkers recently investigated the viability of Jurkat cells by using both a supervised (CNN) and unsupervised (variational autoencoder) machine learning approach [18].

The aim of this study was to evaluate cell stability, i.e., cell viability and membrane integrity, throughout liquid storage of cells in different media. Jurkat cells, an immortalized T cell line, were used as a model for CAR T cells [20], [21]. Two novel AI tools trained on images derived from FIM for assessment of cell viability were applied. In the first approach a CNN was used as image classifier, i.e., the network is intended to process an image of an unknown particle and to assign that particle to one of the four training classes. Additionally, we used the software ParticleSentryAI, which is generating the morphological “fingerprints” of the cells in a similar way as Thite et al. by combining a CNN with a dimensionality reduction to a 2D embedding of the feature vectors. Subsequently, all 2D embeddings of the cells are shown as probability density estimations, i.e., as “fingerprints” using a kernel density estimation [22]. This approach has been successfully applied to characterize differences in morphology of particles in pharmaceutical protein formulations [22], [23].

Within the study, cell viability obtained from traditional flow cytometry with cell staining was compared to viability obtained from FIM by using either cell image classification as well as qualitative morphological “fingerprints”.