(M1530-09-55) Machine Learning-Enabled Morphological Profiling of Primary T Cells Subjected to Forced Degradation for Quality Control Applications

Purpose: Cell-based medicinal products (CBMPs), especially T cell therapies, are transforming treatment options for cancer and other diseases but face substantial manufacturing and analytical challenges owing to their complexity and the fragility of live cells. Traditional assays to assess cell viability and function, such as dye-based staining and flow cytometry, are time-consuming, invasive, and may compromise cell integrity. Recent advances have demonstrated that label-free flow imaging microscopy (FIM), combined with deep learning approaches, can non-invasively assess cell health by extracting and analyzing morphological fingerprints from cell images. Unlike supervised approaches requiring labeled data, we applied an unsupervised variational autoencoder (VAE) with UMAP dimensionality reduction to explore heterogeneous morphologies in primary human T cells exposed to manufacturing-relevant stress conditions. This study aims to demonstrate that unsupervised morphological features from FIM images can distinguish viable, apoptotic, and dead cells and be compared with classical flow cytometry viability and phenotyping metrics. Importantly, this label-free, unsupervised method enables analysis beyond conventional flow cytometry stains and supports near real-time, non-destructive monitoring.

Methods: Primary human T cells were isolated from PBMCs sourced via leukoreduction system (LRS) cones from healthy donors. Following in vitro activation, cells were expanded for 7 days and subjected to acute heat stress and varying cryoprotectant conditions: DMSO concentrations of 0%, 2.5%, 5%, 7.5%, 10%, 20%, and CryoStor 10. Cryopreservation involved cooling to -80°C for 24 hours before transfer to liquid nitrogen storage for one week, simulating standard manufacturing protocols. After thaw, cell suspensions were imaged using FlowCam (FIM), generating high-resolution brightfield images. Morphological features were extracted using an unsupervised VAE followed by UMAP to identify morphologically distinct populations.
Viability and apoptotic status were assessed in parallel using classical flow cytometry: cells were stained with Annexin V and propidium iodide (PI) and measured on a BD LSR II cytometer, with standard compensation and gating strategies to enumerate live (Annexin V-/PI-), apoptotic (Annexin V+/PI-), and dead (Annexin V+/PI+) cells. Activation (CD25, CD69) and exhaustion (PD-1, LAG-3, TIM-3) markers were evaluated before and after stress treatments by multicolor flow cytometry to track functional changes in T cells (Figure 1).

Results: Unsupervised VAE-UMAP analysis effectively segregated viable, apoptotic, and dead T cell populations after heat and cryopreservation stresses (Figure 2). Results from the label-free method are in good agreement with traditional stain-based cytometric viability measurements obtained by using Annexin V and propidium iodide staining. Cell viability was highest at intermediate DMSO concentrations (7.5–10%) and with CryoStor 10, while both lower and higher DMSO concentrations reduced viability (Figure 3). Heat stress elevated apoptotic and dead populations, which also exhibited increased exhaustion marker expression. Activation markers were retained under optimal conditions but diminished under severe stress.

Conclusion: Unsupervised, label-free morphological analysis of FIM images provides a rapid, non-invasive, and robust method to assess primary human T cell health under manufacturing-relevant stresses. This approach aligns well with classical flow cytometry and offers significant potential as a quality control tool, especially for continuous, near real-time monitoring of complex, heterogeneous cell populations beyond the limitations of conventional staining assays.

References: 1. Grabarek, A D et al. “Forced degradation of cell-based medicinal products guided by flow imaging microscopy: Explorative studies with Jurkat cells.” European journal of pharmaceutics and biopharmaceutics: official journal of Arbeitsgemeinschaft fur Pharmazeutische Verfahrenstechnik e.V vol. 167 (2021): 38-47. doi:10.1016/j.ejpb.2021.07.004
2. Thite, Nidhi G et al. “Stain-Free Approach to Determine and Monitor Cell Heath Using Supervised and Unsupervised Image-Based Deep Learning.” Journal of pharmaceutical sciences vol. 113,8 (2024): 2114-2127. doi:10.1016/j.xphs.2024.05.001

Figure 1. Isolation and enrichment of human T cells followed by forced degradation of T cell guided by flow imaging microscopy and flow cytometry.

Figure 2. Density maps showing the 2d feature representations obtained using UMAP for the trained VAE model for images of T cells after heating stress. Regions of high density, indicated by darker shades of green, represent frequently observed cell morphologies, whereas lighter regions correspond to rarer morphological states.

Figure 3. Overlay of flow cytometry density scatter plots of annexin V Alexafluor 488 versus PI for T cells at day 7 before freezing or heating and T cells after different freezing conditions with 0%, 7.5%, 10%, 20%, and CryoStor 10.