Unsupervised Machine Learning-Based Process Analytical Tools for Near Real-Time Cell Morphology Analysis During CAR-T Cell Manufacturing

Cell therapies like Chimeric Antigen Receptor (CAR)-T cell therapy deliver living cells to patients as active pharmaceutical ingredients. Manufacturing of these cells is complex, often yielding, heterogeneous products and high failure rates. Quality control (QC) assays used in CAR-T cell production primarily provide end-point product testing. Real-time process monitoring would be ideal to reduce failure rates and ensure final product quality. However, current analytical tools often fall short due to the heterogeneity of CAR-T cell products and their sensitivity to process changes. In this study, we showcase unsupervised image-based machine learning as a process analytical tool (PAT) for near real-time process monitoring during the production of CAR-T cells. Flow imaging microscopy (FIM) images of T cells collected from nine healthy donors were recorded during the activation, lentiviral-based transduction (expressing CD19 CAR protein), and expansion stages of CAR-T cell production. These images were used to train a Variational Autoencoder (VAE), allowing quantitative tracking of changes in cell morphologies during the various stages of production of CAR-T cells from each donor. Findings include observation of a new, transient population in T cells transduced to express CAR protein. This population was absent in T cells that were not transduced. The density of the new population was proportional to the transduction efficiency determined by traditional stain-based flow cytometry assays. Together, this study demonstrates the utility of using VAEs as a PAT tool for monitoring patient-to-patient variability and early detection of process deviations/upsets.