Background T cell cytotoxicity plays a crucial role in defending against cancer. It is traditionally assessed by endpoint assays with non-microscopic appliance, thus lacking temporal and visual insights into middle time ranges that live-cell imaging offers. However, conventional live-cell imaging falls short in data scale due to limited replicates per condition. Combining micro-structured slides with live-cell imaging addresses these issues. Multiple cell adhesion patterns enable independent progression of killing processes, providing multiplicates of T cell-mediated killing events in one experiment. Handling the high data volume manually can be cumbersome and biased, calling for an objective means of data processing. We were analyzing two cytotoxic T cells with different avidities as determined by IFN-γ release, but failed to clearly differentiate killing capacities using chromium release assay. Therefore, we asked if micro-pattern and AI-assisted analysis could resolve differences in killing.

Materials and Methods Human primary T cells were transduced with T cell receptors of different avidities for the same peptide/MHC complex (T58, D115). D115-T cells were additionally engineered to express chimeric costimulatory switch proteins (CSPs). These CSPs were created by linking the extracellular domain of PD-1 with an intracellular costimulatory domain of CD28 or 4-1BB. Functional T cell avidities were defined by IFN-γ release after co-culture with SkMel23 melanoma cells. Live-cell imaging used micro-patterned µ-Slides 8-Well (ibidi). T cells and tumor cells were used label-free to avoid potential effects on killing processes. Data analysis employed Cellpose, an AI-powered cellular segmentation tool with human-in-the-loop features. Manual quality control and correction, if needed, adopted the interactive tool we built with Python and Jupyter Notebook. R was used to interpret the formatted data.

Results We acquired time series of tumor cell/T cell co-cultures on micro-patterned µ-Slides 8-Well. Cellpose was trained to precisely enclose label-free tumor cell patches formed on the micro-structures. The segmentation was integrated into an automated workflow, whereby the AI automation recognized the outline of tumor cell patches on the micro-structures in sequential time frames and was able to follow each patch over time even if its confluency changed. Applying the live-cell imaging set-up with micro-structured slides together with AI-powered analysis, we were able to resolve different dynamics of confluency changes for high and low avidity T cells indicating different killing behaviors. The automation is now applied to evaluate killing efficacies of T cells with CSP expression and T cells exposed to various components of the tumor microenvironment.

Conclusions Micro-patterns provide standardized cell patches and, hence, high multiplicity within one experiment for statistical robustness. Pattern confluency is reliably measured with trained AI software allowing for processing of high-throughput live-cell imaging data which is essential for selecting T cell variants which best perform in given conditions like the complex tumor microenvironment.

Support JC is funded by China Scholarship Council (No. 202306380076), AF, EN, MS and DR by ZIM/AppMic, and AH by Phio Pharmaceuticals.

J. Cao: None. W. Xu: None. A.J. Fischbeck: None. A.S. Herbstritt: C. Other Research Support (supplies, equipment, receipt of drugs or other in-kind support); Significant; Phio Pharmaceuticals. S. Prins: None. J. Rädler: None. E. Noessner: None. M. Seiwald: A. Employment (full or part-time); Significant; ibidi GmbH. D. Rüdiger: A. Employment (full or part-time); Significant; ibidi GmbH.